Policy parameterization income operation deduction method, apparatus and device, and medium

Abstract

The invention discloses a strategy parameterization income operation deduction method and device, equipment and a medium, and the method comprises the steps: continuously collecting current business operation data and current strategy parameter data in a business execution process, and calculating an income target baseline; generating a predicted income result of the current time point based on the current service operation data and the current strategy parameter data; the predicted income result is compared with an income target baseline in real time, when deviation exists, the contribution degree of each income factor to the deviation is calculated, and a deviation dominant factor is positioned; according to the deviation dominant factors, executing directional disturbance on corresponding strategy parameters in the current strategy parameter data, and generating multiple groups of adjusted strategy combinations; and selecting an optimal strategy combination based on the income deduction result corresponding to each adjusted strategy combination. And through a dynamic deduction mode with a controllable process and an intervened strategy, the initiative, the refinement level and the decision-making efficiency of business revenue management are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of strategy deduction in business scenarios, and in particular to a strategy parameterization revenue operation deduction method, apparatus, equipment and medium. Background Technology

[0002] In recent years, existing revenue management in telecommunications operations has primarily relied on billing systems to perform post-event statistical analysis of transaction data. This post-event statistical model has significant limitations: First, revenue results are delayed, typically only generating complete statistics after the completion of a transaction or billing settlement, making it difficult to promptly identify revenue deviations during operation. Second, the revenue formation process lacks interpretability; when revenue falls short of expectations, it's impossible to quantify whether the cause is business scale, price level, package structure, or channel allocation. Third, there's a lack of predictive capability for strategy adjustments; the effectiveness of different adjustment plans cannot be assessed in advance during business execution, and decisions often rely on experience-based judgment, especially in data-driven services where there's a common dilemma of continuously increasing investment but insufficient revenue growth. In short, current telecommunications operations suffer from problems of delayed and inaccurate operational strategy adjustments. Summary of the Invention

[0003] The purpose of this application is to provide a method, apparatus, device, and medium for parameterizing revenue operation in order to at least solve the problem that current communication operations cannot adjust strategies in real time and efficiently.

[0004] To address the aforementioned technical problems, this application provides a strategy parameterization-based revenue operation extrapolation method, comprising: During business execution, we continuously collect current business operation data and current strategy parameter data, and dynamically calculate the revenue target baseline based on historical data from the same period. Based on the current business operation data and current strategy parameter data, revenue is broken down into a set of quantifiable revenue factors, and the predicted revenue result for the current point in time is generated through an interpretable prediction model. The predicted revenue results are compared with the revenue target baseline in real time. When there is a deviation, the contribution of each revenue factor to the deviation is calculated to identify the dominant factor of the deviation. Based on the deviation-dominant factor, targeted perturbations are performed on the corresponding strategy parameters in the current strategy parameter data to generate multiple sets of adjusted strategy combinations, and the business state transition process under each combination is simulated. Based on the projected revenue results corresponding to each adjusted strategy combination, the optimal strategy combination is selected, converted into a strategy adjustment instruction in natural language format, and pushed to the business execution end to complete the closed-loop intervention.

[0005] Optionally, the dynamic calculation of the revenue target baseline includes: Obtain historical business operation data within a preset time window to extract time-series characteristics of revenue changes; Based on the time-series characteristics of revenue changes, a baseline model is constructed using a sliding window, and the expected revenue range at the current time point is calculated. The expected revenue range at the current time point is calculated based at least on the baseline value obtained from the statistical analysis of historical business operation data within the corresponding sliding window. The baseline weight is dynamically adjusted based on the daily business operation rhythm quantified by the current business operation data, and the expected revenue range at the current time point is dynamically corrected based on the baseline weight to generate the revenue target baseline at the current time point, wherein the baseline is presented in the form of an interval.

[0006] Optionally, the calculation of the contribution of each income factor to the deviation includes: The difference between the predicted revenue result and the revenue target baseline is defined as the total revenue deviation; Obtain the marginal contribution value of each income factor based on the output of the interpretable prediction model, and calculate the contribution ratio of each income factor to the total income deviation based on the marginal contribution value of each income factor. Income factors whose contribution ratio exceeds a preset threshold are marked as deviation-dominant factors.

[0007] Optionally, the simulation of the business state transition process under each combination includes: Construct a state transition model and define the mapping relationship between the current business state vector and the next state vector; Using the adjusted strategy combinations as transition driver inputs, the corresponding state transition paths are calculated respectively. The state transition path is iteratively deduced until the end of the preset time window, generating the cumulative deduction income result corresponding to each combination.

[0008] Optionally, the set of revenue factors includes at least business scale factors, price level factors, package structure factors, channel distribution factors, and time rhythm factors; The interpretable prediction model is a gradient boosting decision tree model or an ensemble learning model, used to output the marginal contribution value of each factor to the prediction result.

[0009] Optionally, the method is applied to communication operation services, wherein performing directional perturbation on the corresponding policy parameters in the current policy parameter data according to the deviation dominance factor includes: When the dominant factor of the deviation is a price factor, a gradient increase or decrease perturbation is applied to the price discount rate in the strategy parameter data. When the dominant deviation factor is a structural factor, a switching disturbance is performed on the main package identifier in the strategy parameter data. When the dominant factor of the deviation is a channel-related factor, a redistribution disturbance is performed on the channel allocation ratio in the strategy parameter data.

[0010] Optionally, the step of selecting the optimal strategy combination based on the inferred revenue results corresponding to each adjusted strategy combination and converting it into a strategy adjustment instruction in natural language format includes: Obtain the policy parameter adjustment amount corresponding to the optimal policy combination; The adjustment amount of the strategy parameters is mapped to a preset instruction template library to generate structured instruction text; The structured instruction text is semantically refined by calling a large language model to generate natural language instructions that are readable by the business.

[0011] To address the aforementioned technical problems, this application also provides a strategy parameterization-based revenue operation simulation apparatus, comprising: The data acquisition module is used to continuously collect current business operation data and current strategy parameter data during business execution, and dynamically calculate the revenue target baseline based on historical data from the same period. The model prediction module is used to decompose revenue into a set of quantifiable revenue factors based on the current business operation data and the current strategy parameter data, and generate the predicted revenue result for the current point in time through an interpretable prediction model; The comparison and inference module is used to compare the predicted income results with the income target baseline in real time. When there is a deviation, it calculates the contribution of each income factor to the deviation and locates the dominant factor of the deviation. The strategy deduction module is used to perform targeted perturbation on the corresponding strategy parameters in the current strategy parameter data according to the deviation dominant factor, generate multiple sets of adjusted strategy combinations, and simulate the business state transition process under each combination. The deduction and optimization module is used to select the optimal strategy combination based on the deduction revenue results corresponding to each adjusted strategy combination, convert it into a strategy adjustment instruction in natural language format, and push it to the business execution end to complete the closed-loop intervention.

[0012] To address the aforementioned technical problems, this application also provides a computer device, including a memory and a processor. The memory stores computer-readable instructions, which, when executed by the processor, cause the processor to perform the steps of the above-described strategy parameterization revenue operation deduction method.

[0013] To address the aforementioned technical problems, this application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the above-described strategy parameterization revenue operation deduction method.

[0014] The beneficial effects of the embodiments of this application are as follows: During business execution, current business operation data and current strategy parameter data are continuously collected, and the revenue target baseline is dynamically calculated based on historical data from the same period. Based on the current business operation data and current strategy parameter data, revenue is decomposed into a set of quantifiable revenue factors, and a predicted revenue result for the current time point is generated through an interpretable prediction model. The predicted revenue result is compared with the revenue target baseline in real time. When a deviation exists, the contribution of each revenue factor to the deviation is calculated, and the dominant factor of the deviation is located. Based on the dominant factor of the deviation, targeted perturbations are performed on the corresponding strategy parameters in the current strategy parameter data to generate multiple sets of adjusted strategy combinations, and the business state transition process under each combination is simulated. Based on the projected revenue results corresponding to each adjusted strategy combination, the optimal strategy combination is selected, converted into a strategy adjustment instruction in natural language format, and pushed to the business execution end to complete closed-loop intervention. By constructing a dynamic revenue target baseline, this solution addresses the lag issue of traditional ex-post statistics, enabling real-time perception of revenue deviations during business execution. Through white-box decomposition of revenue factors and factor-level attribution analysis, revenue deviations are precisely pinpointed to specific dimensions such as scale, price, and structure, resolving the issue of unexplainable revenue formation processes. By using targeted strategy perturbations based on deviation-dominant factors and business state transition simulations, it achieves quantitative pre-evaluation of the effects of different adjustment schemes, overcoming the blindness of strategy adjustments relying on experience-based decisions. Through natural language command generation and closed-loop push notifications from the business execution end, it achieves automated connection from simulation conclusions to actionable interventions. Compared to existing technologies, this solution transforms revenue management from ex-post statistics to a dynamic simulation model with controllable processes and interventionist strategies, significantly improving the initiative, precision, and decision-making efficiency of revenue management. Attached Figure Description

[0015] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A schematic diagram of the basic flow of a strategy parameterization revenue operation extrapolation method according to a specific embodiment of this application; Figure 2 A schematic diagram of the basic structure of a revenue operation simulation device for strategy parameterization according to a specific embodiment of this application; Figure 3 This is a basic structural block diagram of a computer device according to a specific embodiment of this application. Detailed Implementation

[0016] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0017] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the word “comprising” as used in the specification of this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0018] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0019] Those skilled in the art will understand that the term "terminal" as used herein includes both devices that are wireless signal receivers, devices that are wireless signal receivers without transmitting capability, and devices with receiving and transmitting hardware, having receiving and transmitting hardware capable of performing bidirectional communication on a bidirectional communication link. Such devices may include: cellular or other communication devices having a single-line display, a multi-line display, or a cellular or other communication device without a multi-line display; PCS (Personal Communications Service) that can combine voice, data processing, fax, and / or data communication capabilities; PDA (Personal Digital Assistant) that may include a radio frequency receiver, pager, Internet / intranet access, web browser, notepad, calendar, and / or GPS (Global Positioning System) receiver; and conventional laptop and / or handheld computers or other devices that have and / or include a radio frequency receiver. As used herein, "terminal" can be portable, transportable, installed in a means of transportation (air, sea, and / or land), or suitable and / or configured to operate locally, and / or in a distributed manner, operating in any other location on Earth and / or in space. "Terminal" as used herein can also be a communication terminal, an internet access terminal, or a music / video playback terminal, such as a PDA, a MID (Mobile Internet Device), and / or a mobile phone with music / video playback capabilities, or a smart TV, set-top box, etc.

[0020] The hardware referred to by the names "server," "client," and "service node" in this application is essentially an electronic device with the equivalent capabilities of a personal computer. It is a hardware device with the necessary components revealed by the von Neumann architecture, such as a central processing unit (including an arithmetic logic unit and a control unit), memory, input devices, and output devices. The computer program is stored in its memory, and the central processing unit loads the program stored in the secondary storage into the main memory to run it, execute the instructions in the program, and interact with the input and output devices to complete specific functions.

[0021] It should be noted that the concept of "server" used in this application can also be extended to the case of server clusters. Based on the network deployment principles understood by those skilled in the art, the servers should be logically divided. Physically, these servers can be independent of each other but accessible through interfaces, or they can be integrated into a single physical computer or a computer cluster. Those skilled in the art should understand this flexibility and should not use it to constrain the implementation of the network deployment method in this application.

[0022] One or more of the technical features of this application, unless explicitly specified herein, can be deployed on a server and accessed by a client remotely calling the online service interface provided by the server, or can be directly deployed and run on a client for access.

[0023] Unless otherwise specified, the AI ​​models cited or potentially cited in this application may be deployed on a remote server and invoked remotely on the client, or deployed on a client with the capability to invoke directly. In some embodiments, when running on the client, the corresponding intelligence may be acquired through transfer learning in order to reduce the requirements on the client's hardware operating resources and avoid excessive consumption of the client's hardware operating resources.

[0024] Unless otherwise specified, all data involved in this application may be stored remotely on a server or on a local terminal device, as long as it is suitable for use by the technical solution of this application.

[0025] Those skilled in the art will understand that although the various methods in this application are described based on the same concept and thus present commonality among them, they can be performed independently unless otherwise specified. Similarly, the various embodiments disclosed in this application are all based on the same inventive concept; therefore, concepts expressed in the same way, as well as concepts that are appropriately changed for convenience but are expressed differently, should be understood equivalently.

[0026] Unless otherwise expressly stated, the various embodiments disclosed in this application can be combined in a cross-cutting manner to flexibly construct new embodiments, as long as such combination does not depart from the inventive spirit of this application and can meet the needs of the prior art or solve a certain deficiency in the prior art. Those skilled in the art should be aware of such modifications.

[0027] Please see Figure 1 , Figure 1 This is a schematic diagram of the basic process of the revenue operation simulation method for strategy parameterization in this embodiment.

[0028] like Figure 1 As shown, it includes: S1100. During business execution, continuously collect current business operation data and current strategy parameter data, and dynamically calculate the revenue target baseline based on historical data from the same period. This embodiment can be applied to real-time business operation scenarios in various fields such as communications, finance, e-commerce, insurance, education, healthcare, and law, enabling real-time strategy simulation and adjustment for business operations. In this embodiment, an operation system (hereinafter referred to as the "system") equipped with AI models (or AI agents) is configured to dynamically simulate and optimize strategies for various operational scenarios. Specifically, taking business operation strategies in the communications field as an example, during business execution, the system establishes data interfaces with the billing system, customer relationship management system, and strategy configuration platform through a real-time data acquisition module. It continuously acquires current business operation data and current strategy parameter data using a message queue or stream processing framework. The current business operation data includes at least order records, user reach records, channel access records, and traffic usage records; the current strategy parameter data includes at least the main promotion package identifier, price discount rate, channel allocation ratio, and resource investment intensity.

[0029] Based on this, the system performs dynamic revenue target baseline calculation. Specifically, the system first acquires a historical data window corresponding to the current time point, such as business operation data for the same hour over the past 30 days. Then, it uses a sliding window method to extract the temporal characteristics of revenue changes within this window, including the mean, standard deviation, and fluctuation trend. Next, a time weighting factor is introduced, assigning higher weight to recent data to reflect the timeliness of business rhythm. Finally, combined with the business progress already completed that day, a weighted fusion method is used to generate the revenue target baseline for the current time point. This baseline is presented in interval form, including the expected revenue value and its upper and lower fluctuation ranges. Through this method, the revenue target is dynamically updated during business operations, providing a scientific and real-time reference benchmark for subsequent deviation identification.

[0030] It should be noted that the system in this embodiment can use a time-series forecasting model to construct a dynamic baseline. Specifically, the system processes historical business operation data into hourly time series, constructing a triple decomposition model containing trend, periodic, and residual terms. The series is fitted using a seasonal autoregressive moving average model or a Prophet model to predict the expected revenue value for the current time window. During the forecasting process, the system continuously absorbs the business data that has occurred that day as the latest observation of the model, and uses a Bayesian update mechanism to dynamically correct the model parameters, enabling the baseline to quickly respond to changes in the current business rhythm. This approach is suitable for business scenarios with strong periodicity, such as when there is a significant difference in the sales rhythm of packages between weekdays and weekends, and can effectively capture periodic fluctuation patterns.

[0031] It should be noted that this embodiment introduces external features for baseline calibration to improve the accuracy of the dynamic baseline. Specifically, based on historical data from the same period, the system integrates multi-dimensional features such as holiday information, marketing activity information, weather information, and competitor dynamics to construct a feature-enhanced revenue baseline model. The system extracts the encoding vectors of these external factors through feature engineering and inputs them as covariates into the regression model to correct for the offset of the historical baseline. For example, when a large-scale marketing activity is detected on a given day, the system automatically increases the expected revenue baseline; when severe weather is detected that may affect offline channel reach, the system correspondingly lowers the expected revenue baseline. This approach fully considers the impact of external environmental factors on business revenue, making the baseline more closely reflect actual business scenarios.

[0032] It should be noted that, in one implementation, the system employs quantile regression to construct a dynamic threshold range for the income target baseline. Specifically, the system groups historical data for the same period by time point and independently constructs an income distribution model for each time point. Using a quantile regression algorithm, it calculates specified quantiles for the income distribution at each time point, for example, using the 30th quantile as the lower threshold and the 70th quantile as the upper threshold, forming a dynamic baseline range. During actual operation, the system compares the current income progress with the corresponding threshold range. If the income progress is lower than the lower threshold, a negative bias is identified; if it is higher than the upper threshold, a positive bias is identified. This approach can adaptively handle the heteroscedasticity of income distribution at different time points, avoiding false alarms or false negatives caused by using a uniform threshold.

[0033] S1200: Based on the current business operation data and current strategy parameter data, the revenue is decomposed into a set of quantifiable revenue factors, and the predicted revenue result for the current point in time is generated through an interpretable prediction model; During business execution, the system continuously collects current business operation data and current strategy parameter data. After dynamically calculating the revenue target baseline based on historical data from the same period, it decomposes the revenue into a set of quantifiable revenue factors based on the current business operation data and current strategy parameter data. Then, it generates the predicted revenue result for the current point in time through an interpretable prediction model. Specifically, after completing data collection and baseline construction, the system performs white-box decomposition of revenue factors and current prediction. Specifically, the system first performs dimensional modeling on the collected current business operation data and current strategy parameter data to construct a set of quantifiable revenue factors. This set of revenue factors includes at least five categories of quantifiable factors: business scale factors, quantified by statistically analyzing the number of orders placed, the number of users reached, and the number of channel visits within the current time window; price level factors, quantified by calculating the average traffic price, the average package price, and the discount rate; package structure factors, quantified by calculating the sales proportion of each package tier; channel distribution factors, quantified by calculating the order proportion and conversion rate of each channel; and time rhythm factors, quantified by extracting hourly and daily periodic features. After constructing the income factors, the system inputs the feature vectors of these factors into the interpretable prediction model. The interpretable prediction model is preferably a gradient boosting decision tree model. This model iteratively constructs multiple decision trees, each learning the prediction residual of the previous tree. Finally, the prediction results of all trees are summed to generate the predicted income result for the current time point. While outputting the prediction result, the model also outputs the marginal contribution value of each income factor to the prediction result. This marginal contribution value is obtained by calculating the cumulative split gain of each factor in the model's decision path, reflecting the explanatory power of each factor for the current income level. Through this method, the income formation process is transformed from "black box" prediction to "white box" factor decomposition, making the income prediction results traceable and interpretable, providing a quantitative basis for the bias attribution in step 3.

[0034] It should be noted that, in one implementation of this embodiment, the system introduces a feature interaction detection mechanism during the income factor construction process to identify the synergistic effects between factors. Specifically, the system first constructs a basic income factor set, including factors of scale, price, and structure. Subsequently, the system identifies factor pairs with significant interactions by calculating the information gain or interaction strength index between each factor pair, such as the interaction between "channel factor and price factor" or "package structure factor and time factor." The system adds the identified interaction terms as derived factors to the income factor set, forming an enhanced factor feature vector. In the prediction stage, the system uses a deep forest model for prediction. This model extracts feature interaction information layer by layer through a cascaded structure, effectively capturing the complex nonlinear relationships between factors. While outputting the predicted income result, the model outputs the contribution of each factor and interaction term to the prediction result through a feature importance evaluation mechanism, achieving more accurate and interpretable prediction.

[0035] It should be noted that, in another implementation of this embodiment, the system employs a multi-stage decomposition model to achieve white-box prediction of revenue factors, avoiding the use of black-box machine learning models. Specifically, the system decomposes the revenue formation process into three sequential stages: the reach stage, the conversion stage, and the average order value stage. In the reach stage, the system predicts the reach of each channel based on the channel allocation ratio in the current strategy parameter data, combined with historical reach efficiency data. In the conversion stage, the system predicts the conversion rate of each channel based on the price discount rate and the configuration of the main package in the current strategy parameter data, combined with a historical conversion rate model. In the average order value stage, the system predicts the average order value based on the current package sales structure and pricing strategy. The system sums the results of the three stages using a multiplication formula to generate the predicted revenue result for the current point in time. This method is entirely built on business logic, and the calculation process of each stage is transparent and traceable. Simultaneously, the system calculates the sensitivity of each revenue factor to the prediction result using the chain rule, directly outputting the marginal contribution value of each factor for subsequent deviation attribution analysis.

[0036] S1300. The predicted revenue results are compared with the revenue target baseline in real time. When there is a deviation, the contribution of each revenue factor to the deviation is calculated, and the dominant factor of the deviation is located. Based on the current business operation data and current strategy parameter data, revenue is decomposed into a set of quantifiable revenue factors. After generating the predicted revenue result for the current time point through an interpretable prediction model, the predicted revenue result is compared with the revenue target baseline in real time. When a deviation exists, the contribution of each revenue factor to the deviation is calculated to identify the dominant factor of the deviation. Specifically, after generating the predicted revenue result, the system performs deviation detection and factor-level attribution analysis. First, the output predicted revenue result for the current time point is compared with the generated revenue target baseline in real time to calculate the total revenue deviation value. The total revenue deviation value is the difference between the predicted revenue result and the revenue target baseline. When the deviation value exceeds a preset tolerance range, the system determines that there is a revenue deviation. After confirming the existence of a deviation, the system performs factor-level attribution analysis. The system obtains the marginal contribution value of each revenue factor output by the interpretable prediction model. The marginal contribution value reflects the explanatory power of each factor at the current revenue level. The system normalizes the marginal contribution value of each factor and calculates the contribution of each factor to the total revenue deviation. The specific calculation formula is: Factor Contribution = Factor Marginal Contribution Value × Deviation Direction Coefficient, where the deviation direction coefficient is determined by the sign of the deviation and is used to distinguish whether the factor contributes positively or negatively. The system sorts the contributions of each factor from largest to smallest in absolute value and marks them as candidate dominant factors. The system further determines whether the contribution of each candidate dominant factor exceeds a preset threshold. For example, factors with a contribution ratio exceeding 30% are identified as deviation dominant factors. If the contribution of a single factor is insufficient, the system identifies combinations of factors with a cumulative contribution ratio exceeding 60% as deviation dominant factor combinations. Through the above method, precise positioning is achieved from "whether a deviation exists" to "who caused the deviation," providing a clear target for subsequent targeted strategy perturbations.

[0037] It should be noted that, in one implementation of this embodiment, the system uses the Shapley value decomposition method to calculate factor contributions, avoiding dependence on the model's output format. Specifically, when the prediction model cannot directly output the marginal contribution values ​​of each factor, the system uses the Shapley value method for post-hoc attribution. The Shapley value, based on cooperative game theory, calculates the average marginal contribution of each factor to the prediction result across all possible factor combinations. The system simulates the changes in prediction results for each factor under different combinations of values ​​through enumeration or approximate sampling, calculating the Shapley value for each factor. This value satisfies additivity, meaning the sum of the Shapley values ​​of all factors equals the total income deviation. The system uses the Shapley values ​​of each factor as contribution levels, sorts them by absolute value, and identifies the dominant deviation factor. This method is applicable to any type of prediction model, including deep learning black-box models, and has the advantage of model independence.

[0038] It should be noted that, in another implementation, the system employs a decision path backtracking method to locate the dominant bias factor, which is particularly suitable for prediction scenarios based on tree models. Specifically, when using gradient boosting decision trees or random forest models, the system records the complete decision path of the current prediction result on each decision tree in the model. For each decision tree, the system backtracks from the root node to the leaf node, recording the factors used by each split node on the path and their split thresholds. The system statistically analyzes the frequency of factors appearing on all decision trees and calculates the importance score of the factors by combining the weights of each tree. When there is a revenue bias, the system compares the current factor value with the factor distribution during historical normal periods, identifies the factor whose value deviates most significantly from the normal distribution, and marks it as the dominant bias factor. This method not only outputs the factor contribution but also provides a visualized decision path, making it easier for business personnel to understand the attribution logic.

[0039] S1400. Based on the deviation dominance factor, perform directional perturbation on the corresponding strategy parameters in the current strategy parameter data to generate multiple sets of adjusted strategy combinations, and simulate the business state transition process under each combination. The predicted revenue results are compared with the revenue target baseline in real time. When a deviation exists, the contribution of each revenue factor to the deviation is calculated. After identifying the dominant deviation factor, targeted perturbations are performed on the corresponding strategy parameters in the current strategy parameter data based on the dominant deviation factor, generating multiple sets of adjusted strategy combinations, and simulating the business state transition process under each combination. Based on the identification of the dominant deviation factor, targeted perturbations of strategy parameters and state transition simulations are performed. Specifically, the system first matches the corresponding adjustable strategy parameters from the strategy parameter data according to the identified dominant deviation factor type. The mapping relationship between the strategy parameters and revenue factors is pre-configured in the strategy parameter mapping table. For example, when the dominant deviation factor is a price-related factor, the matched strategy parameters are price discount rate and package pricing strategy; when the dominant deviation factor is a structure-related factor, the matched strategy parameter is the main promotion package identifier; when the dominant deviation factor is a channel-related factor, the matched strategy parameter is the channel allocation ratio. The system performs targeted perturbations on the matched strategy parameters to generate multiple sets of adjusted strategy combinations. The directional perturbation employs a gradient-step approach, meaning that based on the current strategy parameter values, adjustments are made gradually in both positive and negative directions at preset step sizes, generating multiple sets of parameter value sequences. For example, regarding price discount rates, the system uses the current discount rate as a benchmark and generates combinations such as a 5% increase, a 10% increase, a 5% decrease, and a 10% decrease. After generating the strategy combinations, the system performs business state transition simulation. The system constructs a state transition matrix, defining the mapping relationship between the current business state vector and the next state vector. The business state vector includes at least the dimensions of current revenue progress, conversion rates of each channel, and sales proportion of each package. The system uses each adjusted strategy combination as the transition driver input and the constructed interpretable predictive model as the state transition function, iteratively simulating the evolution of the business state under each combination along the time axis until the end of the preset simulation time window, outputting the cumulative simulated revenue results corresponding to each combination.

[0040] The above methods enable a complete deduction process from "problem identification" to "strategy generation" and then to "effect simulation".

[0041] It should be noted that, in one implementation, the system uses a Bayesian optimization method to generate the optimal perturbation strategy combination, replacing the ergodic step perturbation. Specifically, the system defines the adjustment space of the strategy parameters as the optimization search domain, using the closeness of the inferred income result to the income target baseline as the objective function. The system uses a Gaussian process as a surrogate model to model the objective function. In each iteration, the system balances exploration and utilization through a collection function, selects the next strategy parameter combination to be evaluated, calls the state transition model for inference and evaluation, and updates the surrogate model based on the evaluation results. After several iterations, the system outputs the strategy parameter combination that makes the inferred income result closest to the target baseline as the optimal perturbation strategy. This approach significantly reduces the computational load compared to ergodic perturbation, and is particularly suitable for complex scenarios with high strategy parameter dimensionality and a large combination space.

[0042] It should be noted that, in one implementation, the system employs a discrete event simulation method to extrapolate business state transitions, replacing continuous state extrapolation based on predictive models. Specifically, the system models the business execution process as a series of discrete events, including user outreach events, order generation events, and package change events. Each event carries a timestamp and business attributes. The system configures the triggering rules of the event generator based on the strategy parameters in the adjusted strategy combination; for example, after adjusting the channel distribution ratio, the frequency of occurrence of outreach events in each channel changes accordingly. The system advances the timeline through an event queue, updating business state variables and accumulating revenue progress each time an event is triggered. After the simulation runs to the end of the extrapolation time window, the system outputs the cumulative extrapolated revenue results for each combination. This method can simulate the randomness and temporal dependencies in the business execution process and is suitable for scenarios requiring fine-grained characterization of micro-level business processes.

[0043] S1500: Based on the projected revenue results corresponding to each adjusted strategy combination, select the optimal strategy combination, convert it into a strategy adjustment instruction in natural language format, and push it to the business execution end to complete the closed-loop intervention.

[0044] After performing targeted perturbations on the corresponding strategy parameters in the current strategy parameter data according to the dominant deviation factor, generating multiple sets of adjusted strategy combinations, and simulating the business state transition process under each combination, the optimal strategy combination is selected based on the projected revenue results corresponding to each adjusted strategy combination. This optimal strategy combination is then converted into a strategy adjustment instruction in natural language format and pushed to the business execution end to complete the closed-loop intervention. Specifically, based on the completion of multi-scenario strategy simulation, the system executes a closed loop of optimal strategy combination selection, instruction generation, and push. The system first ranks and evaluates the projected revenue results corresponding to each adjusted strategy combination. The evaluation indicators include at least the absolute value of the projected revenue result, the degree of closeness to the revenue target baseline, the strategy adjustment cost, and the feasibility of implementation. The system uses a multi-objective decision-making method, such as weighted scoring or analytic hierarchy process, to comprehensively score each strategy combination and select the strategy combination with the highest comprehensive score as the optimal strategy combination. After determining the optimal strategy combination, the system generates natural language instructions. The system pre-builds a strategy instruction template library, which contains instruction templates for different strategy parameter types, such as "It is recommended to adjust the distribution ratio of {channel name} from {original value} to {new value}". The system extracts the adjustment amount of strategy parameters from the optimal strategy combination and fills it into the corresponding instruction template to generate structured instruction text. To further improve the readability and business friendliness of the instructions, the system calls a large language model to semantically refine the structured instruction text, converting it into strategy adjustment instructions in natural language format. Finally, the system pushes the generated strategy adjustment instructions to the business execution end through a message push interface or the API interface of the business execution system. The business execution end includes the operation terminal of the operations personnel or the automated strategy execution engine. The system records the instruction push time, reception status, and execution feedback to form a closed-loop management. If the business execution end returns a confirmation of successful execution, the system will update the strategy parameter data cache to ensure that the strategy parameters used in subsequent simulations are synchronized with the actual execution strategy.

[0045] Through the above methods, a complete business loop is achieved, from "deduction conclusions" to "executable instructions" and then to "execution feedback".

[0046] It should be noted that, in one implementation, the system of this embodiment employs reinforcement learning to dynamically select the optimal strategy, replacing static multi-objective decision-making. Specifically, the system models the business execution process as a Markov decision process, where the state space includes the current business state vector and strategy parameter configuration, the action space consists of the adjustment actions of the strategy parameters, and the reward function is the degree of closeness between the projected revenue result and the revenue target baseline. The system trains the strategy selection model using a deep Q-network or a policy gradient method, which can output the optimal strategy action in real time based on the current business state. Each time a deviation triggers a projection, the system calls the trained strategy selection model, inputs the current business state, and directly outputs the optimal strategy combination, without needing to traverse multiple combinations for projection. This approach enables real-time response in strategy selection and continuous optimization of the selected strategy as business data accumulates.

[0047] Furthermore, the system in this embodiment employs a large language model to implement an interactive instruction generation and confirmation mechanism, enhancing the understandability of strategy adjustment instructions and the ability for manual intervention. Specifically, the system combines contextual information such as the optimal strategy combination, projected revenue results, deviation attribution information, and historical strategy execution records into structured prompts, which are then input into the large language model. Based on understanding the aforementioned context, the large language model generates natural language instruction text containing the reasons for strategy adjustment, expected effects, and specific operational steps. The instruction text adopts a conversational style, for example, "The current revenue deviation is mainly caused by the excessive proportion of low-priced packages. It is recommended to switch the main package from package A to package B, which is expected to make up for about 70% of the revenue gap. Do you want to execute?" The system pushes the generated instruction to the operator's terminal, supporting manual confirmation, rejection, or execution after adjustment. This approach improves the transparency and controllability of strategy intervention and is suitable for high-risk business scenarios or those requiring careful manual decision-making.

[0048] This solution can be applied to data-driven business scenarios for telecommunications operators. For example, during a monthly mid-month promotional campaign for data packages, a provincial branch can collect real-time data on outbound call reach, app exposure, order volume for each data package tier, and current discount strategies. Based on historical revenue trends, a daily revenue target baseline is dynamically constructed. When the system predicts that the daily cumulative revenue is 15% below the baseline, it identifies the deviation as primarily caused by an excessively high proportion of low-value packages through factor-level attribution. It then performs targeted perturbations on the strategy parameters, simulating a combination of switching the main package from the 19 yuan tier to the 39 yuan tier and simultaneously increasing the outbound call channel allocation by 10%. The simulation results show that this adjustment can make up approximately 80% of the revenue shortfall within the remaining time. The system pushes the above strategy adjustment instructions to the operations management terminal in natural language. After confirmation by operations personnel, the system can execute the changes with a single click, transforming revenue management from a post-event statistical model to a process-driven model.

[0049] In the above implementation, by constructing a dynamic revenue target baseline, the lag problem of traditional ex-post statistics is solved, enabling real-time perception of revenue deviations during business execution. Through white-box decomposition of revenue factors and factor-level attribution analysis, revenue deviations are precisely located to specific dimensions such as scale, price, and structure, resolving the issue of the inexplicable revenue formation process. By using targeted strategy perturbations based on deviation-dominant factors and business state transition simulations, quantitative effect pre-evaluation of different adjustment schemes is achieved, resolving the problem of blind reliance on experience-based decision-making in strategy adjustments. Through natural language command generation and closed-loop push from the business execution end, automated connection from simulation conclusions to executable interventions is achieved. Compared to existing technologies, this solution transforms revenue management from ex-post statistics to a dynamic simulation model with controllable processes and interventionist strategies, significantly improving the initiative, precision, and decision-making efficiency of revenue management.

[0050] In some implementations, the dynamic calculation of the revenue target baseline includes: S1111. Obtain historical business operation data within a preset time window to extract time-series characteristics of revenue changes; In this embodiment, during the dynamic calculation of the revenue target baseline, historical business operation data within a preset time window is first acquired to extract the time-series characteristics of revenue changes. Specifically, the system acquires historical business operation data within the preset time window and extracts the time-series characteristics of revenue changes from this data. Specifically, the system first determines the time attributes of the current point in time, including hour, day of the week, and whether it is a holiday. Based on these time attributes, the system retrieves historical business operation data with the same time attributes as the current point in time from the historical database. For example, if the current time is Wednesday at 2 PM, the system retrieves all business operation data for Wednesdays at 2 PM within the past N weeks. The preset time window can be configured according to the characteristics of the business cycle, preferably the past 30 days or the past 8 weeks. After acquiring the historical data, the system extracts the time-series characteristics of revenue changes. The system arranges the revenue data for each historical period in chronological order to construct a revenue time-series sequence. Based on this sequence, the system extracts the following time-series features: mean, reflecting the average income level of the same period in history; standard deviation, reflecting the degree of income fluctuation; trend slope, calculated by linear regression fitting to determine the direction and rate of income change over time; periodicity strength, calculated by autocorrelation function to determine the periodic stability of the income sequence; and quantile features, including the 25th, 50th, and 75th quantiles, used to describe the income distribution pattern. These time-series features serve as the basic input for subsequent baseline calculations, used to generate a dynamic income target baseline that reflects historical patterns. Through this method, the quantitative extraction of historical business patterns is achieved, providing data support for the scientific construction of dynamic baselines.

[0051] It should be noted that, in one implementation, the system employs a time-series decomposition method to extract the periodic characteristics of revenue changes. Specifically, the system acquires historical revenue time-series data within a preset time window and uses the STL decomposition algorithm to decompose the time-series data into a trend term, a seasonal term, and a residual term. The trend term reflects the long-term direction of revenue changes; the seasonal term reflects the periodic fluctuation pattern of revenue, including daily, weekly, and monthly cycles; and the residual term reflects the random fluctuation component. The system extracts the seasonal index for each time point in the seasonal term to quantify the degree of deviation of revenue from the average level in different time periods. For example, the system can calculate the seasonal index at 2 PM on Wednesday; if the index is greater than 1, it indicates that the revenue level at that time is usually higher than the average level for the whole day. The system outputs the latest slope of the trend term, the seasonal index for each time point in the seasonal term, and the standard deviation of the residual term as time-series features. This method can more precisely depict the multiple periodic patterns of revenue and is suitable for scenarios with complex business rhythms and overlapping multiple cycles.

[0052] In another alternative implementation, the system employs a business rhythm grouping method to extract weighted time-series features, enhancing the timeliness of recent data. Specifically, the system first groups historical data from the same period according to business execution rhythm, for example, processing weekday and weekend data separately, or promotional and non-promotional days separately. For each group of data, the system further assigns different weight coefficients according to time proximity, using an exponential decay function to calculate the weights, with historical data closer to the current time having higher weights. Based on the weighted data, the system calculates the weighted mean, weighted standard deviation, and weighted trend slope. The weighted mean is obtained by summing the products of each period's revenue value and its weight, divided by the total weights; the weighted trend slope is obtained by fitting using weighted least squares, with recent data having a greater impact on the slope. This approach enables the extracted time-series features to more sensitively reflect recent business trends, making it suitable for scenarios where business patterns evolve over time.

[0053] S1112. Based on the time-series characteristics of revenue changes, a baseline model is constructed using a sliding window, and the expected revenue range at the current time point is calculated. The expected revenue range at the current time point is calculated at least based on the baseline value obtained from the statistical analysis of historical business operation data within the corresponding sliding window. Furthermore, after acquiring historical business operation data within a preset time window to extract revenue change time-series features, a sliding window baseline model is constructed based on these features, and the expected revenue range for the current time point is calculated. Specifically, the system uses a sliding window method to construct the baseline model. The sliding window refers to a data window that maintains a fixed length on the time axis and slides forward over time. The system uses the revenue change time-series features extracted in the previous steps as samples within the window. The window length can be configured according to business fluctuation characteristics, preferably the most recent 7 days or the most recent 14 days. The system calculates a statistical baseline value based on the historical revenue data within the window. First, the weighted mean of the revenue data within the window is calculated as the expected revenue value. The weight allocation uses an exponential decay method, with higher weights for data closer to the current time within the window. Second, the weighted standard deviation of the revenue data within the window is calculated as a volatility measure. The system constructs the expected revenue range based on the weighted mean and weighted standard deviation. The expected revenue range is represented as [expected revenue value - k × weighted standard deviation, expected revenue value + k × weighted standard deviation], where k is the confidence coefficient, which can be configured according to business tolerance, preferably 1.96 corresponding to a 95% confidence level. The sliding window is dynamically updated as the current time step progresses. Each time the time step moves forward by one unit, the system removes the oldest data from the window and adds the latest data to the window, recalculating the weighted mean and weighted standard deviation within the window to generate the expected revenue range for the current time point. This method achieves dynamic updating of the baseline, ensuring that the expected revenue range can promptly reflect recent business trends.

[0054] It should be noted that, in one implementation, the system in this embodiment uses an adaptive sliding window method to construct the baseline model, with the window length dynamically adjusted according to business volatility. Specifically, the system first calculates the coefficient of variation (COP) of the data within the window, i.e., the ratio of the standard deviation to the mean, to measure the relative volatility of the data. If the COP exceeds a preset threshold, it indicates significant business volatility, and the system automatically reduces the window length to allow the baseline to respond more quickly to recent changes. If the COP is below the preset threshold, it indicates relatively stable business, and the system automatically expands the window length to include more historical samples to enhance stability. The system further employs a dynamic time warp algorithm to evaluate the similarity between the data within the window and the current business rhythm. When the similarity is below the threshold, samples with low similarity are actively removed, retaining only historical data consistent with the current rhythm for baseline calculation. This approach can adaptively handle changes in business volatility and is suitable for scenarios with unstable business rhythms or sudden disturbances.

[0055] In an alternative implementation, the system employs quantile regression to construct robust expected income intervals, replacing the interval construction method based on mean and standard deviation. Specifically, the system uses extracted time-series features as independent variables and historical income values ​​as dependent variables to construct multiple quantile regression models, including 10th, 25th, 75th, and 90th quantile models. Each quantile model is fitted independently, outputting the corresponding quantile's income estimate. The system uses the interval between the 25th and 75th quantiles as the expected income interval and the interval between the 10th and 90th quantiles as the warning interval. This approach does not assume that income data follows a normal distribution, and can more robustly handle scenarios where the income distribution is skewed or contains outliers. Furthermore, the quantile interval width can adaptively reflect the fluctuation characteristics of income at different levels; for example, the interval width automatically expands during income peaks and automatically narrows during income troughs.

[0056] S1113. The baseline weight is dynamically adjusted based on the daily business operation rhythm quantified by the current business operation data, and the expected revenue range at the current time point is dynamically corrected based on the baseline weight to generate the revenue target baseline at the current time point, wherein the baseline is presented in the form of an interval.

[0057] Furthermore, a sliding window baseline model is constructed based on the aforementioned time-series characteristics to calculate the expected revenue range at the current time. The baseline weights are dynamically adjusted according to the daily business operation rhythm to generate the revenue target baseline. Based on the completed sliding window baseline calculation, the baseline weights are dynamically adjusted according to the daily business operation rhythm to generate the final revenue target baseline. Specifically, the system first acquires the actual business operation data that has occurred from the current day to the present time, including the number of orders generated, the amount of revenue realized, and the percentage of time progress completed. The system quantifies the daily business operation rhythm coefficient by calculating the ratio of actual revenue progress to time progress, using the formula: Rhythm Coefficient = Actual Revenue Progress ÷ Time Progress. If the rhythm coefficient is greater than 1, it indicates that the business operation is faster than the normal rhythm; if it is less than 1, it indicates that the business operation is slower than the normal rhythm. The system uses the rhythm coefficient as a weight adjustment factor to dynamically correct the generated expected revenue range. Specifically, the system multiplies the expected revenue value by the rhythm coefficient to generate the rhythm-calibrated expected revenue value. Simultaneously, the system adjusts the range width according to the deviation of the rhythm coefficient. When the rhythm coefficient deviates significantly from 1, the range width is appropriately increased to accommodate greater uncertainty. The adjustment formula is: calibrated interval width = original interval width × (1 + α × |rhythm coefficient - 1|), where α is a preset amplification coefficient. The system further introduces a time decay factor to assign weights to the actual revenue data that occurred that day, and then weights and fuses it with the rhythm-calibrated expected revenue value. As time progresses, the weight of the actual revenue data gradually increases, while the weight of the predicted component gradually decreases. Finally, a revenue target baseline for the current time point is generated. This baseline is presented in interval form, including the calibrated expected revenue value and its upper and lower fluctuation ranges.

[0058] Through the above methods, the baseline is adaptively adjusted to the real-time business rhythm of the day, making the revenue target baseline closer to the actual business operation status.

[0059] In one implementation, the system uses a piecewise linear interpolation method to dynamically adjust the rhythm weights, replacing overall calibration based on a single rhythm coefficient. Specifically, the system divides the daily business operation timeline into multiple time periods, such as morning, afternoon, and evening peak hours, and calculates the rhythm coefficient independently for each period. The system constructs a piecewise linear interpolation function for each period, using the time points already occurred within the period as independent variables and the actual revenue progress as the dependent variable, to interpolate and calculate the corresponding value of the current time point on the expected revenue curve for the day. The system then concatenates the interpolation results from each period to form a complete daily rhythm calibration curve. During baseline fusion, the system uses differentiated fusion weights for different time periods. For example, for completed periods, the actual revenue data has a 100% weight; for the current period, the actual revenue data and predicted data are weighted proportionally according to their time progress; and for future periods, the system relies entirely on predicted data. This approach can more precisely characterize the differentiated features of business rhythm across different time periods and is suitable for business scenarios with significant fluctuations in daily revenue rhythm.

[0060] In an alternative implementation, the system introduces a real-time anomaly detection mechanism to identify and correct the impact of abnormal events during the dynamic adjustment of baseline weights. Specifically, while calculating the daily business operation rhythm coefficient, the system monitors for the presence of anomaly event markers, including system failures, sudden marketing campaigns, and sudden changes in the external environment. The system uses an anomaly detection model to identify whether current business data deviates from the normal pattern. This anomaly detection model can employ an isolated forest or autoencoder method, trained based on historical normal data. When an anomaly event is detected, the system matches the generated expected revenue range with the anomaly event type, queries a pre-set anomaly event correction library, and obtains the correction coefficient for the impact of that type of event on revenue. The system then adds the correction coefficient to the rhythm calibration process to generate a revenue target baseline that considers the impact of anomaly events. This approach avoids baseline distortion caused by anomalies and improves the robustness of the baseline.

[0061] This embodiment achieves refined and dynamic construction of the revenue target baseline by introducing historical data from the same period to extract time-series features, constructing a sliding window baseline model, and dynamically adjusting weights according to the daily business operation rhythm. Compared with static target settings, this technical solution can adaptively reflect the periodic patterns and intraday rhythm changes of the business, making the revenue target baseline more scientific, real-time, and in line with the actual business operation status. It provides an accurate and reliable reference benchmark for subsequent deviation detection, effectively reducing the risk of false alarms or missed alarms caused by unreasonable baseline settings.

[0062] In some implementations, the calculation of the contribution of each income factor to the deviation in S1300 includes: S1311. The difference between the predicted revenue result and the revenue target baseline is determined as the total revenue deviation; During the deviation detection process, the difference between the predicted revenue result and the revenue target baseline is determined as the total revenue deviation. Specifically, the system first obtains the output predicted revenue result at the current time point and the generated revenue target baseline at the current time point. The revenue target baseline is presented in the form of an interval, including the expected revenue value, a lower threshold, and an upper threshold. When calculating the total revenue deviation, the system adopts a segmented judgment mechanism. If the predicted revenue result is between the lower threshold and the upper threshold of the revenue target baseline, it is determined that there is no deviation, and the total revenue deviation value is zero. If the predicted revenue result is lower than the lower threshold, the difference between the predicted revenue result and the lower threshold is calculated as a negative deviation value; if the predicted revenue result is higher than the upper threshold, the difference between the predicted revenue result and the upper threshold is calculated as a positive deviation value. The formula for calculating the deviation value is: Negative deviation value = lower limit threshold - predicted revenue result (when predicted revenue result < lower limit threshold) Positive deviation value = Predicted revenue result - Upper limit threshold (when predicted revenue result > upper limit threshold) The system stores the calculated deviation value as the total income deviation variable and simultaneously records the deviation direction identifier, where negative deviation is identified by -1, positive deviation by +1, and no deviation by 0. This deviation direction identifier is used to determine the contribution direction of each factor in subsequent factor-level attribution analysis. This method achieves a quantitative definition of income deviation, providing a clear deviation input for subsequent contribution calculations.

[0063] It should be noted that, in one extended implementation, the system employs a dynamic threshold mechanism to determine multi-level deviations, rather than a single binary judgment. Specifically, based on the revenue target baseline, the system further constructs multi-level early warning threshold ranges. These multi-level thresholds include a monitoring threshold, an early warning threshold, and an alarm threshold. Each threshold is dynamically set based on the quantiles of the historical deviation distribution; for example, the 75th quantile of historical deviation values ​​is set as the monitoring threshold, the 90th quantile as the early warning threshold, and the 95th quantile as the alarm threshold. The system compares the difference between the predicted revenue result and the revenue target baseline level by level with each threshold to determine the deviation level. When the absolute value of the deviation exceeds the monitoring threshold but does not exceed the early warning threshold, it is marked as a monitoring-level deviation; when it exceeds the early warning threshold but does not exceed the alarm threshold, it is marked as an early warning-level deviation; and when it exceeds the alarm threshold, it is marked as an alarm-level deviation. The system outputs the total revenue deviation value along with the deviation level identifier, which is used to trigger differentiated processing strategies in subsequent steps. This approach enables tiered responses based on the severity of the deviation, avoiding excessive intervention in minor fluctuations.

[0064] In an alternative implementation, the system uses a time-series cumulative method to calculate the cumulative deviation within a sliding window, replacing the deviation calculation for single-point times. Specifically, the system constructs a fixed-length sliding window, the window length of which can be configured to several past time units, such as the past 6 hours or the past 24 hours. The system calculates the deviation between the predicted revenue result and the revenue target baseline at each time point within the window, and calculates the cumulative deviation using a weighted summation method. The weight allocation adopts an exponential decay method, with deviation points closer to the current time having higher weights. The formula for calculating the cumulative deviation is: Cumulative Deviation = Σ(Deviation Value_i × Weight_i), where i iterates through each time point within the window. The system outputs the cumulative deviation as the total revenue deviation to trigger subsequent attribution and extrapolation processes. This method can smooth out random fluctuations at single-point times, avoid false triggers due to accidental factors, and is suitable for scenarios where business data contains a certain amount of noise.

[0065] S1312. Obtain the marginal contribution value of each income factor based on the output of the interpretable prediction model, and calculate the contribution ratio of each income factor to the total income deviation based on the marginal contribution value of each income factor. In this embodiment, after determining the difference between the predicted income result and the income target baseline as the total income deviation, the contribution ratio of each income factor to the total income deviation is calculated based on the marginal contribution values ​​of each factor output by the interpretable prediction model. Specifically, after determining the total income deviation, the contribution ratio of each income factor to the total income deviation is calculated based on the marginal contribution values ​​of each factor output by the interpretable prediction model. Specifically, the system first obtains the marginal contribution values ​​of each income factor output by the interpretable prediction model. These marginal contribution values ​​reflect the influence of each factor on the predicted income result at the current value level, and are usually presented in the form of feature importance scores or SHAP values. The system further obtains the target reference value corresponding to each income factor during the income target baseline construction process. The target reference value can be the mean, median, or weighted average of the factor in the same historical period, and the specific value selection method is consistent with the baseline construction method. The system calculates the deviation between the current value and the target reference value of each factor, using the formula: Deviation = |Current Value - Target Reference Value| / Target Reference Value, used to quantify the degree to which the factor actually deviates from the normal level. The system multiplies the marginal contribution value of each factor with its deviation to obtain the original contribution of that factor to the total income deviation. The economic meaning of this product is: the factor's influence on the prediction result multiplied by the actual degree of deviation from the normal level determines the magnitude of that factor's contribution to the total income deviation. The system normalizes the original contributions of all factors and calculates the contribution percentage of each factor using the formula: Contribution Percentage_i = Original Contribution_i / Σ(Original Contribution_j). The sum of the contribution percentages is 100%, reflecting the relative importance of each factor in the total income deviation. The system stores the calculated contribution percentages of each factor along with the corresponding factor name, current value, target reference value, and deviation direction as input data for step 3.3, which is then used to identify the dominant factor.

[0066] The above methods enable the quantitative transformation from factor influence to deviation contribution, providing a quantitative basis for accurately locating the source of deviation.

[0067] S1313, Mark the income factors whose contribution ratio exceeds a preset threshold as deviation-dominant factors.

[0068] This embodiment calculates the contribution percentage of each revenue factor to the total revenue deviation based on the marginal contribution values ​​of each factor output by the interpretable prediction model. Factors with contribution percentages exceeding a preset threshold are then marked as dominant deviation factors. Specifically, after calculating the contribution percentage of each revenue factor, the system performs the identification and marking of dominant deviation factors. Specifically, the system first obtains a list of contribution percentages for each revenue factor, sorted from largest to smallest, and includes the factor name, contribution percentage value, and deviation direction identifier. The system presets a dominant factor identification threshold, which can be configured according to the business scenario, preferably 30%. The system traverses the sorted factor list, sequentially determining whether the contribution percentage of each factor exceeds the preset threshold. If a factor has a contribution percentage exceeding 30%, the system marks it as a dominant deviation factor and records its deviation direction, for example, "price factor (negative deviation, contribution percentage 45%)". If the contribution percentage of any single factor does not exceed the preset threshold, the system uses a cumulative contribution method to identify the combination of dominant factors. The system starts with the factor with the largest contribution percentage and sequentially accumulates the contribution percentages of factors until the cumulative contribution exceeds a preset cumulative threshold. The cumulative threshold is preferably 60%. The system marks the set of factors with a cumulative contribution exceeding 60% as the dominant factor combination of deviation, and outputs them in descending order of contribution percentage. For example, if the contribution percentage of "structural factors" is 25%, "price factors" is 22%, and "scale factors" is 18%, and their cumulative contribution reaches 65%, the system marks these three factors together as the dominant factor combination of deviation. After marking, the system outputs the information of the dominant factor of deviation. This information includes the type of dominant factor, contribution percentage, deviation direction, and the degree of deviation between the current factor value and the target reference value. For multi-factor combinations, the system also outputs prompts on the interaction relationships between the factors, allowing for comprehensive consideration of the synergistic effects between factors during subsequent strategy perturbations. This achieves the function of accurately identifying core problem factors from the contribution of multiple factors, providing a clear target for targeted strategy perturbations.

[0069] This implementation method defines total income deviation as the difference between the predicted result and the baseline. It calculates the contribution ratio based on the factor marginal contribution value output by the interpretable model and the factor deviation degree, and employs a threshold screening mechanism to accurately identify the dominant factor of deviation, achieving a quantitative decomposition of income deviation from the overall to the factor level. This technical solution solves the problems of traditional attribution analysis relying on experience-based judgment and having vague localization, making the source of deviation quantifiable and traceable. This provides a clear and precise target for subsequent targeted strategy perturbations, significantly improving the efficiency and accuracy of problem localization.

[0070] In some implementations, S1400 simulates the service state transition process under each combination, including: S1411. Construct a state transition model and define the mapping relationship between the current business state vector and the next state vector; In this embodiment, during strategy deduction, a state transition model is constructed to define the mapping relationship between the current business state vector and the next state vector. Specifically, the system first defines the dimensional composition of the business state vector. The business state vector includes at least the following dimensions: current cumulative revenue, current revenue progress, conversion rate of each channel, sales proportion of each package, current strategy parameter configuration, time progress, and remaining time window. Each dimension is represented numerically to form a fixed-dimensional state vector. The system constructs a state transition model, which describes the evolution of the business state to the next time point under the influence of the current state and current strategy parameters. The state transition model is constructed by extending an interpretable prediction model into a state transition function. Specifically, the system uses the current state vector and the strategy parameters to be executed as the model input, and the predicted state vector at the next time point as the model output. The state transition function is expressed as: S_{t+1} = F(S_t, P_t), where S_t is the current state vector, P_t is the current strategy parameter combination, and F is the state transition function. During the model training phase, the system utilizes historical business operation data, taking the actual state vectors at each time point as samples, the current state and current policy parameters as input features, and the next state vector as the prediction target to train the state transition model. The trained model can predict the evolution trend of business states based on the current state and policy parameters. This method constructs a state transition model capable of simulating the dynamic evolution of business processes, providing a core computing engine for subsequent policy deduction.

[0071] It should be noted that, in one implementation, the system employs a Hidden Markov Model (HMM) to construct state transition relationships, suitable for scenarios where business states contain unobservable variables. Specifically, the system divides business states into observable states and latent states. Observable states include directly collectable indicators such as cumulative revenue and conversion rates; latent states include factors that are difficult to observe directly, such as user preferences and market competitiveness. The system constructs an HMM, defining the transition probability matrix between latent states and the emission probability distribution from latent states to observable states. Model training uses the Baum-Welch algorithm, estimating model parameters based on historical observation sequences. During the extrapolation process, the system infers the probability distribution of the current latent state based on the current observable state, and then predicts the probability distribution of the latent and observable states at the next moment. This approach can capture the impact of potential factors that are difficult to observe directly on state transitions during business processes.

[0072] In an alternative implementation, the system employs a graph neural network to construct a relational state transition model, suitable for scenarios where complex interactions exist between different dimensions of the business state. Specifically, the system constructs each dimension of the state vector as graph nodes and the correlations between dimensions as graph edges, forming a state relationship graph. The correlations can be determined using Pearson correlation coefficients or mutual information from historical data. The system uses a graph convolutional network or a graph attention network as the state transition model, with the current state vector and its relationship graph structure as input, and the next state vector as output. The graph neural network, through a neighbor node information aggregation mechanism, can automatically learn the interactive influences between different state dimensions; for example, how changes in channel conversion rates affect the percentage of package sales through the relationship graph. This approach can more accurately model the complex collaborative relationships between business state dimensions and is suitable for highly coupled, multi-dimensional business scenarios.

[0073] S1412. Using the adjusted strategy combinations as transition driver inputs, calculate the corresponding state transition paths respectively; Furthermore, after constructing the state transition model and defining the mapping relationship between the current business state vector and the next state vector, the adjusted strategy combinations are used as transition driving inputs to calculate the corresponding state transition paths. Specifically, the system first obtains the generated M sets of adjusted strategy combinations, denoted as {P1, P2, …, P_M}, and the output current time point business state vector S_current. For each strategy combination P_i, the system performs iterative calculation of the state transition path. The system inputs the current state vector S_current and the strategy combination P_i into the state transition model constructed in step 4.1, and the model outputs the state vector S_{t+1} at the next time point. The system uses S_{t+1} as the new current state and inputs it again with the same strategy combination P_i into the state transition model to iteratively generate the state vectors at subsequent time points. The number of iterations is determined by the preset projection time window length. If the projection window is T time units, then iterates T times to generate the state sequence [S1, S2, …, S_T]. During the iteration process, the system records key state indicators at each time point, including cumulative revenue, conversion rates of each channel, and sales share of each package, forming a complete state transition path. For different strategy combinations, the system executes the above iterative calculations in parallel, generating M independent state transition paths. A convergence judgment mechanism is introduced into the calculation process. If the change in the state vector at multiple consecutive time points is less than a preset threshold, the system determines that the business state is stabilizing and terminates the iteration early to improve deduction efficiency. The system uses the cumulative revenue value in the final state vector of each path as the deduced revenue result for that strategy combination, while retaining the complete intermediate states of the path for subsequent analysis and visualization. This method achieves quantitative simulation of the business evolution trajectory under multiple strategy combinations, providing a data foundation for subsequent strategy effectiveness evaluation.

[0074] S1413. Iterate along the state transition path to the end of the preset time window and generate the cumulative deduction income result corresponding to each combination.

[0075] After using each adjusted strategy combination as a transition driver input and calculating the corresponding state transition path, the system iteratively extrapolates along the transition path to the end of a preset time window, generating the cumulative extrapolated revenue result for each combination. Specifically, the system first obtains the generated state transition path sequence, where each path contains the state vector for each time step from the current point in time to the end of the preset extrapolated time window. Then, the system iteratively extrapolates along the time axis until the end of the preset time window is reached. The preset time window end can be configured according to business management needs, such as the end of the day, the end of the week, or the end of the month. In each iteration, the system extracts the new revenue value generated within the current time step from the current state vector and adds it to the cumulative revenue variable of the strategy combination. The formula for calculating cumulative revenue is: Cumulative Revenue = Σ(New Revenue at Each Time Step), where the new revenue is calculated by the difference in cumulative revenue between adjacent time steps in the state vector. During the iterative extrapolation process, the system synchronously records key intermediate indicators for each time step, including cumulative revenue progress, the trajectory of conversion rate changes for each channel, and the evolution trend of sales proportion for each package. These intermediate indicators, along with the cumulative revenue results, are stored together to form a complete deduction record for subsequent strategy performance comparison and analysis. After all strategy combinations have completed iterative deduction, the system outputs a list of cumulative deduction revenue results for each combination. For each strategy combination, the system also outputs its deduction confidence index, calculated based on the prediction error distribution of the state transition model on historical data, reflecting the reliability of the deduction results. The system sorts the cumulative deduction revenue results from highest to lowest and passes them along with the corresponding strategy combination, deduction path, and confidence information to subsequent steps for optimal strategy selection. This completes the quantitative output of multi-strategy combination deduction results, providing data support for strategy optimization.

[0076] This implementation method defines the mapping relationship of business state vectors by constructing a state transition model. It uses strategy combinations as transition-driven inputs to calculate the corresponding state transition paths and iteratively extrapolates along these paths to the end of the time window, generating cumulative extrapolated revenue results. This achieves quantitative simulation of the business evolution process and parallel evaluation of the effects of multiple strategies. This technical solution solves the problems of inability to pre-calculate strategy adjustments and difficulty in quantifying their effects in existing technologies. It enables decision-makers to predict the revenue impact of different adjustment schemes in a virtual environment, significantly improving the scientific nature and efficiency of strategy decision-making.

[0077] In some implementations, the set of revenue factors includes at least business scale factors, price level factors, package structure factors, channel distribution factors, and time rhythm factors; the interpretable prediction model is a gradient boosting decision tree model or an ensemble learning model, used to output the marginal contribution value of each factor to the prediction result.

[0078] In this embodiment, the revenue factor set includes specific types and the specific implementation methods of the interpretable prediction model. Specifically, the system first constructs quantitative calculation methods for five types of revenue factors. For business scale factors, the system calculates the total number of orders, active users, and total channel reach within the current time window. For price level factors, the system calculates the average traffic price (total traffic revenue divided by total traffic usage), the average package price (total package revenue divided by total orders), and the discount level (the ratio of actual revenue to standard revenue). For package structure factors, the system calculates the sales proportion of each package tier (e.g., low-end, mid-range, high-end) and constructs a structure concentration index, such as the Herfindahl-Hirschman Index, to quantify the concentration of package distribution. For channel distribution factors, the system calculates the order proportion and conversion rate of each channel (e.g., APP, outbound calls, business halls) and calculates the channel diversity index. For time rhythm factors, the system extracts hourly, daily, and weekly periodic characteristics, including the cumulative proportion of current revenue to total daily revenue.

[0079] After factor construction, the system employs a gradient boosting decision tree model as the interpretable prediction model. Specifically, the system uses the XGBoost or LightGBM framework, taking the five types of factors mentioned above as feature inputs and using the current point-in-time income as the prediction target for model training. After model training, the system outputs the marginal contribution value of each factor to the prediction result through a feature importance evaluation mechanism. This marginal contribution value is obtained by calculating the cumulative split gain of each factor in the model's decision path, reflecting the explanatory power of that factor for income changes. For ensemble learning models, the system can use a random forest model, calculating feature importance based on the average reduction in out-of-bag data error.

[0080] It should be noted that, in one implementation, the system further constructs factor interaction terms based on the five basic factors to capture the synergistic effects between factors. Specifically, the system identifies factor pairs with significant interactions through statistical analysis methods, such as the interaction between channel factors and price factors, or the interaction between package structure factors and time rhythm factors. The system constructs interactive factors using a product approach, such as the product of channel conversion rate and discount rate, to quantify the cumulative effect when high discounts are offered through high-conversion channels. The system also incorporates interactive factors into the revenue factor set, enabling the model to learn the nonlinear synergistic relationships between factors. Regarding the interpretable prediction model, the system employs a deep forest model, which automatically extracts high-order feature interactions through a cascaded structure while maintaining interpretability. This approach is suitable for business scenarios with complex interactions between factors and can improve prediction accuracy.

[0081] It should be noted that, to suit scenarios with strong multicollinearity among factors, the system employs a linear regression model combined with ridge regression to achieve interpretable predictions. Specifically, the system standardizes the five types of income factors and constructs a multiple linear regression model, with the predicted income value being the weighted sum of all factors. Due to the potential for high correlation among factors, the system introduces ridge regression for parameter estimation, using L2 regularization to constrain the coefficient magnitude and reduce the impact of multicollinearity on the stability of coefficient estimation. The regression coefficients of each factor output by the model are directly used as marginal contribution values, their economic meaning being the change in income caused by a unit change in this factor, assuming other factors remain constant. This approach offers the strongest interpretability, as the coefficients directly reflect the marginal contribution of the factors, and it boasts high computational efficiency, making it suitable for business scenarios with relatively stable revenue formation mechanisms.

[0082] This embodiment decomposes revenue into five quantifiable factors: business scale, price level, package structure, channel distribution, and time rhythm. It then uses a gradient boosting decision tree or ensemble learning model to output the marginal contribution value of each factor, achieving a "white-box" decomposition and interpretable prediction of the revenue formation process. This technical solution solves the problems of traditional revenue forecasting models being black-box and difficult to trace the causes of revenue changes. It enables business personnel to clearly identify the degree of influence of each factor on revenue, providing a transparent and quantifiable decision-making basis for subsequent deviation attribution and strategy adjustment.

[0083] In some implementations, the method is applied to communication operation services, and in S1400, directional perturbation is performed based on the deviation dominance factor, including: S1421: When the dominant factor of the deviation is a price factor, perform gradient increase or decrease perturbation on the price discount rate in the strategy parameter data; In this embodiment, when the dominant deviation factors are of different types, the system performs differentiated directional perturbation operations. Technically, the system first matches the corresponding adjustable policy parameters from the policy parameter data based on the output dominant deviation factor type, and then executes the corresponding perturbation logic.

[0084] Specifically, when the dominant factor of deviation is a price-related factor, the system performs gradient increment / decrement perturbations on the price discount rate in the strategy parameter data. Specifically, the system obtains the currently executed price discount rate, denoted as d_current. Using d_current as a baseline, the system generates perturbation sequences in both positive and negative directions according to a preset step size (preferably 2% or 5%). Positive perturbations generate combinations such as d_current + step and d_current + 2×step to simulate price increases; negative perturbations generate combinations such as d_current - step and d_current - 2×step to simulate price decreases or increased discounts. The system sets perturbation boundaries to ensure that the price discount rate does not exceed preset upper thresholds (e.g., 100%) and lower thresholds (e.g., 50%), avoiding the generation of unreasonable strategy combinations.

[0085] It should be noted that, in one implementation, this method is applied to dynamic pricing scenarios on e-commerce platforms. When the dominant factor of deviation is a price-related factor, the system executes an adaptive perturbation strategy based on price elasticity, rather than a simple gradient increase or decrease. Specifically, the system first calculates the price elasticity coefficient of the current product, i.e., the percentage change in demand when the price changes by 1%. If the product is highly elastic, the system uses small-step perturbations to avoid drastic fluctuations in sales caused by large price adjustments; if the product is inelastic, the system uses large-step perturbations to compensate for revenue shortfalls through more significant price adjustments. The system further introduces a competitor price monitoring mechanism; when competitor prices change, the system automatically adjusts the perturbation boundaries to generate a price strategy combination with market competitiveness. This approach is suitable for price-sensitive business scenarios such as e-commerce and retail.

[0086] In another extended application, this method is used in the marketing and recommendation of financial products. When the dominant deviation factor is a structural factor, the system executes a structured product recommendation perturbation based on user segmentation, rather than simply switching the main promotion package. Specifically, the system first divides the user group into multiple segmented customer groups, such as high-net-worth customers, price-sensitive customers, and product-preference customers. Based on the historical conversion characteristics of each customer group, the system generates differentiated product recommendation combinations; for example, recommending high-value products to high-net-worth customers and discounted products to price-sensitive customers. The system achieves targeted optimization of the product structure by adjusting the weight distribution of the recommendation strategy for each customer group. The system uses a multi-armed slot machine algorithm to explore online within the perturbation combination and dynamically adjusts the recommendation strategy based on real-time feedback. This approach is suitable for financial business scenarios with rich product structures, such as bank wealth management and insurance.

[0087] S1422: When the dominant deviation factor is a structural factor, perform a switching disturbance on the main package identifier in the strategy parameter data; In this embodiment, when the dominant deviation factor is a structural factor, the system executes a switching perturbation on the main promotion package identifier to optimize the package sales structure. Specifically, the system first obtains the current main promotion package identifier, denoted as P_current, and the alternative package list, denoted as P_alt = [P1, P2, …, P_k], from the strategy parameter data. The alternative package list is predefined according to the business strategy and usually includes package types that are substitutes for the current main promotion package, such as different tiers of data packages or different price-level bundled packages. The system generates multiple switching perturbation combinations, each switching the main promotion package identifier to a package in the alternative package list. For each switching combination, the system keeps other strategy parameters (such as price discount rate and channel allocation ratio) unchanged, constructing a single-variable perturbation scheme. The system further generates composite perturbation combinations, linking the main promotion package switching with price discount rate adjustments. For example, when the main promotion package switches from a low-end package to a high-end package, the system simultaneously generates a perturbation combination with a moderately increased price discount rate to simulate the accompanying pricing strategy when promoting a high-end package.

[0088] When executing switching perturbations, the system incorporates a package compatibility verification mechanism. If the alternative packages conflict with the current channel delivery strategy—for example, if a package is only available through a specific channel—the system automatically filters the switching combinations for that package, ensuring that the generated strategy combinations are business-compliant. The system also sets perturbation boundaries, limiting the number of packages switched simultaneously and the magnitude of the switch, avoiding the generation of overly aggressive strategy combinations. Through these methods, the system can generate multiple sets of structure optimization strategy solutions, providing rich strategy inputs for subsequent simulations.

[0089] S1423: When the dominant factor of the deviation is a channel-related factor, a redistribution disturbance is performed on the channel allocation ratio in the strategy parameter data.

[0090] In this embodiment, when the dominant deviation factor is a channel-related factor, the system performs a reallocation perturbation on the channel allocation ratio in the strategy parameter data to optimize channel resource allocation. Specifically, the system first obtains the current allocation ratio vector for each channel, denoted as C_current = [c1, c2, …, c n [ ], where n is the total number of channels, and each component satisfies Σc_i = 1. The channels include APP push notifications, outbound marketing calls, SMS push notifications, offline service halls, and third-party cooperative channels. The system generates multiple channel allocation schemes. The system first obtains historical conversion rate data and marginal benefit curves for each channel. The marginal benefit curves reflect the increasing or decreasing trend of conversion rates when the channel allocation ratio increases. Based on the marginal benefit of each channel, the system uses a ratio adjustment algorithm to generate perturbation combinations.

[0091] The system executes three types of redistribution perturbations: The first is single-channel adjustment, selecting the channel with the highest conversion rate and increasing its allocation by a preset step size, while simultaneously reducing the allocation of other channels proportionally, generating a new allocation vector. The second is inter-channel transfer, transferring a portion of the allocation from low-conversion channels to high-conversion channels, with the transfer magnitude generated according to a step size gradient, such as 5%, 10%, or 15%. The third is multi-dimensional optimization, performing Pareto optimal allocation based on the marginal benefits of each channel, generating multiple optimal allocation combinations by solving a constrained optimization problem. When generating redistribution schemes, the system introduces channel capacity constraints to ensure that the adjusted allocation ratio does not exceed the maximum capacity of each channel. For example, the outbound call channel is limited by the number of agents, and its allocation ratio cannot exceed the upper limit threshold. The system also sets adjustment step size boundaries to avoid excessively large single adjustments that could impact business operations. The system outputs the generated multiple redistribution allocation ratio vectors as channel-type perturbation combinations for subsequent state transition deduction. Through these methods, the system can generate diverse channel optimization schemes, providing strategic support for improving overall channel conversion efficiency.

[0092] This implementation method employs differentiated, targeted perturbation strategies based on the type of dominant deviation factor. When the dominant deviation factor is a price-related factor, it applies gradient increases or decreases to the price discount rate; when it is a structural factor, it applies a perturbation to switch the main promotion package identifier; and when it is a channel-related factor, it applies a perturbation to redistribute the channel allocation ratio. This achieves precise positioning and differentiated generation of strategy adjustments. This technical solution solves the problems of traditional "one-size-fits-all" and lack of specificity in strategy adjustments, ensuring that strategy perturbations are precisely matched to the root cause of the problem. This significantly improves the efficiency and effectiveness of strategy derivation and reduces the trial-and-error costs associated with ineffective strategy attempts.

[0093] In some implementations, S1500 generates policy adjustment instructions in natural language format, including: S1511. Obtain the policy parameter adjustment amount corresponding to the optimal policy combination; In this embodiment, during the strategy instruction generation process, the system first obtains the strategy parameter adjustment amount corresponding to the optimal strategy combination. Specifically, the system first obtains the output list of each adjusted strategy combination and its corresponding projected revenue results. The list contains M strategy combinations, each containing complete strategy parameter configurations and the cumulative revenue results generated by the project. Then, the system performs the optimal strategy combination selection operation. The system uses a multi-objective decision-making method to comprehensively score the strategy combinations. The multi-objective decision-making comprehensively considers the following indicators: the closeness between the projected revenue results and the revenue target baseline, calculated as revenue closeness = 1 - |projected revenue - target revenue| / target revenue; strategy adjustment cost, calculated based on the parameter adjustment magnitude and adjustment type; implementation feasibility score, scored based on whether the strategy combination conflicts with existing business rules; and strategy stability score, assessing the variance of revenue fluctuations after strategy adjustment. After standardizing each indicator, the system calculates the comprehensive score of each strategy combination using a weighted summation method, with weights configured according to business priorities. The system selects the strategy combination with the highest comprehensive score as the optimal strategy combination. After obtaining the optimal strategy combination, the system compares it item by item with the currently executed strategy parameter data to extract the strategy parameter adjustment amount. The adjustment amount includes the parameter type, adjustment direction, adjustment magnitude, and values ​​before and after the adjustment. For example, for the price discount rate, the adjustment amount is recorded as "Price discount rate, increase, +5%, from 80% to 85%"; for the main promotion package identifier, the adjustment amount is recorded as "Main promotion package identifier, switch, switch from package A to package B"; for the channel allocation ratio, the adjustment amount is recorded as a vector of ratio changes for each channel. The system stores the extracted strategy parameter adjustment amounts in a structured format as input data for generating instructions in subsequent steps. This method enables the scientific selection of the optimal solution from multiple strategy combinations and the quantification of the adjustment magnitude.

[0094] It should be noted that, in one implementation, the system employs the Pareto front method for optimal strategy selection, suitable for scenarios where multiple objectives conflict. Specifically, the system uses multiple evaluation metrics of strategy combinations as dimensions in the objective space, constructing M distribution points of strategy combinations in the objective space. The system identifies the non-dominated solution set, i.e., the Pareto front, through Pareto dominance relations. A non-dominated solution is defined as one where no other solution is inferior to it on all objectives and superior to it on at least one objective. The system marks all strategy combinations on the Pareto front as candidate optimal solutions, rather than forcibly selecting one. The system provides an interactive selection interface based on different business preferences, allowing decision-makers to select the final execution strategy on the Pareto front according to priority. This approach preserves multiple optimal possibilities and is suitable for scenarios where business objectives are diverse and difficult to quantify uniformly.

[0095] In an alternative implementation, the system employs a robust optimization method to select robust strategies to address prediction uncertainty. Specifically, the system not only uses the single projected revenue result but also incorporates the prediction error distribution of the state transition model over historical data. The system performs multiple Monte Carlo simulations for each strategy combination, adding noise by randomly sampling from the error distribution in each simulation to generate multiple possible outcome scenarios. The system calculates the revenue distribution for each strategy combination across these scenarios, including the mean, standard deviation, and worst-case revenue value. The system selects the strategy combination that minimizes the worst-case loss, i.e., the robust strategy that maximizes the worst-case revenue. This approach reduces the impact of prediction model errors on strategy selection and is suitable for business scenarios with low risk tolerance.

[0096] S1512. Map the adjustment amount of the strategy parameters to a preset instruction template library to generate structured instruction text.

[0097] After obtaining the adjustment amount of the strategy parameters corresponding to the optimal strategy combination, the adjustment amount is mapped to a preset instruction template library to generate structured instruction text. Specifically, the system pre-builds a strategy instruction template library, which contains structured instruction templates for different strategy parameter types. For example, the price-related template is "Change {parameter name} from {original value} {adjustment direction} {adjustment amount}"; the package switching template is "Change the main package from {original package} to {new package}"; and the channel-related template is "Adjust the distribution ratio of {channel name} from {original ratio} to {new ratio}". The system matches the corresponding instruction template from the template library according to the parameter type in the strategy parameter adjustment amount. The system fills the specific value in the adjustment amount into the slot of the template to generate structured instruction text. The system simultaneously generates combined text of multiple instructions to form a complete set of strategy adjustment instructions. This achieves standardized conversion from quantitative adjustment amounts to structured instructions, providing standardized input for subsequent natural language processing.

[0098] S1513. Call the large language model to semantically refine the structured instruction text and generate natural language instructions that are readable by the business.

[0099] After mapping the strategy parameter adjustment amounts to a preset instruction template library to generate structured instruction text, the system calls a large language model to semantically refine the structured instruction text, generating business-readable natural language instructions. Specifically, the system first constructs prompt words, organizing the structured instruction text, the reasons for the strategy adjustment, and the projected revenue results as input. An example prompt word is: "Convert the following strategy adjustment instruction into business-friendly, concise, and clear natural language, and explain the reasons for the adjustment: Increase the price discount rate from 80% to 85%, which is expected to make up for a revenue shortfall of approximately 1.2 million yuan." The system calls the large language model interface, submitting the prompt words to the model for reasoning. Based on its language understanding and generation capabilities, the large language model converts the structured instruction into a fluent natural language expression, such as "It is recommended to increase the current package discount rate from 80% to 85%, which is expected to make up for a revenue shortfall of approximately 1.2 million yuan. Please confirm the execution." The system outputs the refined natural language instruction and pushes it to the business execution end. This enhances the readability of the instructions, improving the acceptance and execution efficiency of operations personnel.

[0100] This implementation method obtains the adjustment amount of strategy parameters for the optimal strategy combination, maps it to a preset instruction template library to generate structured instruction text, and calls a large language model to perform semantic polishing to generate business-readable natural language instructions. This achieves automated transformation from quantitative deduction results to executable strategy instructions. This technical solution solves the problem that traditional strategy output only stays at the data level and is difficult for business personnel to directly understand and execute. It significantly improves the readability and operability of instructions, reduces the cost of manual interpretation, and forms a complete business closed loop from deviation identification to strategy execution.

[0101] Please refer to details. Figure 2 , Figure 2 This is a schematic diagram of the basic structure of the revenue operation simulation device for strategy parameterization in this embodiment.

[0102] like Figure 2As shown, a revenue operation simulation device with parameterized strategy includes: a data acquisition module 1100, used to continuously collect current business operation data and current strategy parameter data during business execution, and dynamically calculate the revenue target baseline based on historical data from the same period; a model prediction module 1200, used to decompose revenue into a set of quantifiable revenue factors based on the current business operation data and current strategy parameter data, and generate the predicted revenue result for the current time point through an interpretable prediction model; a comparison and inference module 1300, used to compare the predicted revenue result with the revenue target baseline in real time, and when there is a deviation, calculate the contribution of each revenue factor to the deviation and locate the dominant factor of the deviation; a strategy simulation module 1400, used to perform directional perturbation on the corresponding strategy parameters in the current strategy parameter data according to the dominant factor of the deviation, generate multiple sets of adjusted strategy combinations, and simulate the business state transition process under each combination; and a simulation optimization module 1500, used to select the optimal strategy combination based on the simulation revenue result corresponding to each adjusted strategy combination, convert it into a strategy adjustment instruction in natural language format, and push it to the business execution end to complete the closed-loop intervention.

[0103] The aforementioned strategy-parameterized revenue performance simulation device continuously collects current business operation data and current strategy parameter data during business execution, and dynamically calculates the revenue target baseline based on historical data from the same period. Based on the current business operation data and current strategy parameter data, revenue is decomposed into a set of quantifiable revenue factors, and a predictive revenue result for the current time point is generated using an interpretable prediction model. The predicted revenue result is compared with the revenue target baseline in real time. When a deviation exists, the contribution of each revenue factor to the deviation is calculated, and the dominant deviation factor is identified. Based on the dominant deviation factor, targeted perturbations are applied to the corresponding strategy parameters in the current strategy parameter data to generate multiple adjusted strategy combinations, and the business state transition process under each combination is simulated. Based on the simulated revenue results corresponding to each adjusted strategy combination, the optimal strategy combination is selected, converted into a strategy adjustment instruction in natural language format, and pushed to the business execution end to complete closed-loop intervention. By constructing a dynamic revenue target baseline, this solution addresses the lag issue of traditional ex-post statistics, enabling real-time perception of revenue deviations during business execution. Through white-box decomposition of revenue factors and factor-level attribution analysis, revenue deviations are precisely pinpointed to specific dimensions such as scale, price, and structure, resolving the issue of unexplainable revenue formation processes. By using targeted strategy perturbations based on deviation-dominant factors and business state transition simulations, it achieves quantitative pre-evaluation of the effects of different adjustment schemes, overcoming the blindness of strategy adjustments relying on experience-based decisions. Through natural language command generation and closed-loop push notifications from the business execution end, it achieves automated connection from simulation conclusions to actionable interventions. Compared to existing technologies, this solution transforms revenue management from ex-post statistics to a dynamic simulation model with controllable processes and interventionist strategies, significantly improving the initiative, precision, and decision-making efficiency of revenue management.

[0104] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed]. Figure 3 , Figure 3 This is a basic structural block diagram of the computer device in this embodiment.

[0105] like Figure 3 The diagram shows the internal structure of a computer device. This computer device includes a processor, non-volatile storage medium, memory, and a network interface connected via a system bus. The non-volatile storage medium stores the operating system, database, and computer-readable instructions; the database may store control information sequences. The processor provides computational and control capabilities, supporting the operation of the entire computer device. The memory stores computer-readable instructions, which, when executed by the processor, cause the processor to execute a policy-parameterized revenue calculation method. The network interface is used for communication with terminals. Those skilled in the art will understand that… Figure 3The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0106] In this embodiment, the processor is used to execute... Figure 2 The system comprises a data acquisition module 1100, a model prediction module 1200, a comparison and inference module 1300, a strategy deduction module 1400, and a deduction optimization module 1500. During business execution, it continuously collects current business operation data and current strategy parameter data, and dynamically calculates the revenue target baseline based on historical data from the same period. Based on the current business operation data and current strategy parameter data, revenue is decomposed into a set of quantifiable revenue factors, and a predicted revenue result for the current time point is generated through an interpretable prediction model. The predicted revenue result is compared with the revenue target baseline in real time. When a deviation exists, the contribution of each revenue factor to the deviation is calculated, and the dominant deviation factor is identified. Based on the dominant deviation factor, targeted perturbations are applied to the corresponding strategy parameters in the current strategy parameter data to generate multiple adjusted strategy combinations, and the business state transition process under each combination is simulated. Based on the deduced revenue results corresponding to each adjusted strategy combination, the optimal strategy combination is selected, converted into a strategy adjustment instruction in natural language format, and pushed to the business execution end to complete closed-loop intervention. By constructing a dynamic revenue target baseline, this solution addresses the lag issue of traditional ex-post statistics, enabling real-time perception of revenue deviations during business execution. Through white-box decomposition of revenue factors and factor-level attribution analysis, revenue deviations are precisely pinpointed to specific dimensions such as scale, price, and structure, resolving the issue of unexplainable revenue formation processes. By using targeted strategy perturbations based on deviation-dominant factors and business state transition simulations, it achieves quantitative pre-evaluation of the effects of different adjustment schemes, overcoming the blindness of strategy adjustments relying on experience-based decisions. Through natural language command generation and closed-loop push notifications from the business execution end, it achieves automated connection from simulation conclusions to actionable interventions. Compared to existing technologies, this solution transforms revenue management from ex-post statistics to a dynamic simulation model with controllable processes and interventionist strategies, significantly improving the initiative, precision, and decision-making efficiency of revenue management.

[0107] This application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the revenue operation deduction method for policy parameterization described in any of the above embodiments.

[0108] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0109] Those skilled in the art will understand that the steps, measures, and solutions in the various operations, methods, and processes discussed in this application can be alternated, modified, combined, or deleted. Furthermore, other steps, measures, and solutions in the various operations, methods, and processes discussed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted. Furthermore, steps, measures, and solutions in the prior art that are similar to those disclosed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted.

[0110] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A strategy-parameterized revenue operation extrapolation method, characterized in that, include: During business execution, we continuously collect current business operation data and current strategy parameter data, and dynamically calculate the revenue target baseline based on historical data from the same period. Based on the current business operation data and current strategy parameter data, revenue is broken down into a set of quantifiable revenue factors, and the predicted revenue result for the current point in time is generated through an interpretable prediction model. The predicted revenue results are compared with the revenue target baseline in real time. When there is a deviation, the contribution of each revenue factor to the deviation is calculated to identify the dominant factor of the deviation. Based on the deviation-dominant factor, targeted perturbations are performed on the corresponding strategy parameters in the current strategy parameter data to generate multiple sets of adjusted strategy combinations, and the business state transition process under each combination is simulated. Based on the projected revenue results corresponding to each adjusted strategy combination, the optimal strategy combination is selected, converted into a strategy adjustment instruction in natural language format, and pushed to the business execution end to complete the closed-loop intervention.

2. The revenue operation extrapolation method with strategy parameterization according to claim 1, characterized in that, The dynamically calculated revenue target baseline includes: Obtain historical business operation data within a preset time window to extract time-series characteristics of revenue changes; Based on the time-series characteristics of revenue changes, a baseline model is constructed using a sliding window, and the expected revenue range at the current time point is calculated. The expected revenue range at the current time point is calculated based at least on the baseline value obtained from the statistical analysis of historical business operation data within the corresponding sliding window. The baseline weight is dynamically adjusted based on the daily business operation rhythm quantified by the current business operation data, and the expected revenue range at the current time point is dynamically corrected based on the baseline weight to generate the revenue target baseline at the current time point, wherein the baseline is presented in the form of an interval.

3. The revenue operation extrapolation method with strategy parameterization according to claim 1, characterized in that, The calculation of the contribution of each income factor to the deviation includes: The difference between the predicted revenue result and the revenue target baseline is defined as the total revenue deviation; Obtain the marginal contribution value of each income factor based on the output of the interpretable prediction model, and calculate the contribution ratio of each income factor to the total income deviation based on the marginal contribution value of each income factor. Income factors whose contribution ratio exceeds a preset threshold are marked as deviation-dominant factors.

4. The revenue operation extrapolation method with strategy parameterization according to claim 1, characterized in that, The simulated business state transition process under each combination includes: Construct a state transition model and define the mapping relationship between the current business state vector and the next state vector; Using the adjusted strategy combinations as transition driver inputs, the corresponding state transition paths are calculated respectively. The state transition path is iteratively deduced until the end of the preset time window, generating the cumulative deduction income result corresponding to each combination.

5. The revenue operation extrapolation method with strategy parameterization according to claim 1, characterized in that, The set of revenue factors includes at least business scale factors, price level factors, package structure factors, channel distribution factors, and time rhythm factors; The interpretable prediction model is a gradient boosting decision tree model or an ensemble learning model, used to output the marginal contribution value of each factor to the prediction result.

6. The revenue operation extrapolation method with strategy parameterization according to claim 5, characterized in that, The method is applied to communication operation services. The step of performing targeted perturbation on the corresponding strategy parameters in the current strategy parameter data based on the deviation dominance factor includes: When the dominant factor of the deviation is a price factor, a gradient increase or decrease perturbation is applied to the price discount rate in the strategy parameter data. When the dominant deviation factor is a structural factor, a switching disturbance is performed on the main package identifier in the strategy parameter data. When the dominant factor of the deviation is a channel-related factor, a redistribution disturbance is performed on the channel allocation ratio in the strategy parameter data.

7. The revenue operation extrapolation method with strategy parameterization according to claim 1, characterized in that, The process of selecting the optimal strategy combination based on the inferred revenue results corresponding to each adjusted strategy combination and converting it into a strategy adjustment instruction in natural language format includes: Obtain the policy parameter adjustment amount corresponding to the optimal policy combination; The adjustment amount of the strategy parameters is mapped to a preset instruction template library to generate structured instruction text; The structured instruction text is semantically refined by calling a large language model to generate natural language instructions that are readable by the business.

8. A revenue operation simulation device with strategy parameterization, characterized in that, include: The data acquisition module is used to continuously collect current business operation data and current strategy parameter data during business execution, and dynamically calculate the revenue target baseline based on historical data from the same period. The model prediction module is used to decompose revenue into a set of quantifiable revenue factors based on the current business operation data and the current strategy parameter data, and generate the predicted revenue result for the current point in time through an interpretable prediction model; The comparison and inference module is used to compare the predicted income results with the income target baseline in real time. When there is a deviation, it calculates the contribution of each income factor to the deviation and locates the dominant factor of the deviation. The strategy deduction module is used to perform targeted perturbation on the corresponding strategy parameters in the current strategy parameter data according to the deviation dominant factor, generate multiple sets of adjusted strategy combinations, and simulate the business state transition process under each combination. The deduction and optimization module is used to select the optimal strategy combination based on the deduction revenue results corresponding to each adjusted strategy combination, convert it into a strategy adjustment instruction in natural language format, and push it to the business execution end to complete the closed-loop intervention.

9. A computer device, characterized in that, The system includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, cause the processor to perform the steps of the income run deduction method for policy parameterization as described in any one of claims 1 to 7.

10. A storage medium storing computer-readable instructions, characterized in that, When the computer-readable instructions are executed by one or more processors, the one or more processors perform the steps of the revenue run-out method for policy parameterization as described in any one of claims 1 to 7.

Images