Short message marketing effect optimization method and system based on multi-channel dynamic weight

Abstract

A short message marketing effect optimization method and system based on multi-channel dynamic weight, the method steps are as follows: obtaining a plurality of short message channel real-time performance data, obtaining standardized performance index through Z-score standardization processing; calculating the channel performance benchmark score through the weighted sum formula; collecting multi-dimensional performance data in the past 24 hours, analyzing the historical conversion data time sequence characteristics by using a three-layer LSTM network, and predicting the conversion rate trend change of each channel in the future 2-4 hours; combining the conversion rate trend change and the channel performance benchmark score, taking the overall ROI maximization as the target, solving the weight distribution scheme by using the gradient optimization method, and calculating the optimal distribution ratio; adjusting the short message queue shunting strategy in real time according to the optimal distribution ratio, starting the weight configuration smooth transition mechanism, and realizing dynamic optimization. The application realizes dynamic optimization of short message marketing effect.

Description

Technical Field

[0001] This invention relates to the field of SMS marketing technology, and in particular to a method and system for optimizing SMS marketing effectiveness based on multi-channel dynamic weights, which is used to achieve intelligent allocation and effectiveness optimization of SMS marketing channel resources. Background Technology

[0002] SMS marketing, as an important component of digital marketing, plays a crucial role in areas such as financial services, e-commerce promotion, and customer outreach. With the deepening development of the mobile internet, businesses are increasingly demanding higher precision and conversion rates from their SMS marketing efforts.

[0003] Currently, SMS marketing systems generally employ a multi-channel parallel sending strategy, which involves simultaneously sending marketing SMS messages to the target user group through the interfaces of multiple SMS service providers. Common technical solutions include two modes: polling distribution and fixed weight allocation. Polling distribution uses each channel in a preset order, while fixed weight allocation sets the sending ratio of each channel based on historical experience.

[0004] The most typical existing technology is a multi-channel management system based on static weight configuration. This system achieves batch distribution of SMS messages by pre-setting the sending ratio of each channel (e.g., channel A occupies 60% of the traffic, channel B occupies 40% of the traffic) and combining it with reach rate monitoring. Its technical principle is to divide the queue of SMS messages to be sent according to a preset ratio, route them to the corresponding channel interfaces, and monitor the SMS delivery status of each channel. When the reach rate of a certain channel falls below a threshold, manual intervention is triggered for adjustment.

[0005] However, this statically configured technical solution suffers from significant response lag and a single evaluation dimension. The system cannot dynamically adjust resource allocation strategies based on real-time conversion results, leading to inefficient channels continuously consuming valuable sending resources, while also lacking comprehensive consideration of key business indicators such as conversion rate and complaint rate. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for optimizing SMS marketing effectiveness based on multi-channel dynamic weights. This method can standardize the real-time performance data of multiple SMS channels, calculate the comprehensive performance score of each channel by weighted summation, predict the future conversion rate trend of each channel, calculate the optimal distribution ratio of each channel by gradient optimization method, and adjust the SMS queue diversion strategy in real time, thereby achieving dynamic optimization of SMS marketing effectiveness.

[0007] To achieve the above objectives, the technical solution provided by the present invention is as follows:

[0008] A method for optimizing SMS marketing effectiveness based on multi-channel dynamic weighting includes:

[0009] Obtain real-time performance data for multiple SMS channels;

[0010] The real-time performance data is standardized by using the Z-score standardization method to map reach rate, conversion rate, complaint rate, and latency index to the [0,1] interval to obtain standardized performance indexes.

[0011] Based on the standardized performance indicators, the overall performance score of each channel is calculated by weighted summation, using the following scoring formula:

[0012] Generate a channel performance benchmark score, with reach rate as a weight. Conversion rate weight Complaint rate weighting Delay index weight satisfy Establish a weight allocation scheme;

[0013] Collect multi-dimensional performance data of each channel over the past 24 hours, analyze the temporal characteristics of historical conversion data through LSTM neural network, and use a three-layer LSTM network structure for temporal pattern learning to predict the trend of conversion rate changes of each channel in the next 2-4 hours.

[0014] Based on the conversion rate trend and the channel performance benchmark score, a gradient optimization method is used to solve for a solution that maximizes the overall ROI while satisfying the constraints. The weight allocation scheme, wherein the gradient optimization method solves for the objective function:

[0015]

[0016] Where n is the number of channels. This represents the weight ratio assigned to the i-th SMS channel, where the predicted conversion rate i represents the average conversion rate of the i-th SMS channel over the next 2-4 hours, and the weight coefficient i represents the adjustment parameter related to the channel characteristics in the i-th SMS channel. i =Basic coefficient i × quality factor i ×Cost Factors i The quality factor i The base coefficient is calculated based on the channel performance benchmark score of the i-th SMS channel. i The cost factor is used to reflect the importance of the i-th SMS channel. i It is inversely proportional to the sending cost of the i-th SMS channel;

[0017] Based on the objective function, the Adam optimization algorithm is used to solve it, with a learning rate of 0.01 and a maximum number of iterations of 100, to calculate the weight allocation vector that converges to the optimal solution;

[0018] Obtain the weight allocation vector, compare and analyze it with the current weight configuration, and output the optimal distribution ratio;

[0019] Based on the optimal distribution ratio, the SMS queue diversion strategy is adjusted in real time, and the SMS messages to be sent are redistributed to each channel queue. At the same time, a smooth transition mechanism for weight configuration is activated to achieve dynamic optimization of SMS marketing effectiveness.

[0020] Preferably, the standardization of the real-time performance data involves using the Z-score standardization method to map reach rate, conversion rate, complaint rate, and latency indicators to the [0,1] interval, resulting in standardized performance indicators, including:

[0021] Collect raw data on reach rate, conversion rate, complaint rate, and delay indicators for each SMS channel, calculate the mean μ and standard deviation σ of each indicator in historical data, and establish a data distribution baseline;

[0022] Based on the aforementioned data distribution baseline, the Z-score normalization formula is applied. Where X represents the original data value;

[0023] Normalization transformation is performed on each indicator to eliminate the influence of differences in the dimensions of different indicators, and the normalization transformation result is obtained.

[0024] The normalization transformation result is subjected to interval mapping, and the Z value is mapped to the standard interval [0,1] using the Sigmoid function to obtain the standardized performance index.

[0025] Preferably, the step of collecting multi-dimensional performance data of each channel over the past 24 hours, analyzing the temporal characteristics of historical conversion data through an LSTM neural network, and using a three-layer LSTM network structure for temporal pattern learning to predict the conversion rate trend of each channel in the next 2-4 hours includes:

[0026] Based on the multi-dimensional performance data, the moving average of the conversion rate over 3 hours, 6 hours, and 24 hours is calculated, and trend features such as the first-order difference and second-order difference of the conversion rate are constructed to form the input feature vector of the prediction model.

[0027] Based on the input feature vector of the prediction model, a three-layer LSTM network structure is designed, wherein the input layer receives feature vectors of 24 time windows, and the hidden layer captures long-short-term dependencies through 128 LSTM units to establish a time series prediction model.

[0028] The time-series prediction model is deployed as an online inference service, and combined with the standardized performance indicators, the conversion rate trend of each channel is predicted in the next 2-4 hours.

[0029] Preferably, the smooth transition mechanism for simultaneously activating weight configuration includes:

[0030] To obtain the optimal distribution ratio, an exponential decay smoothing algorithm is used, with the smoothing coefficient α set to 0.3, to establish a weight smoothing transition formula.

[0031] Based on the weight smoothing transition formula, calculate the smoothed weight = α × updated weight + (1-α) × current weight, and generate a progressive weight adjustment sequence;

[0032] The weight switching operation is performed step by step according to the progressive weight adjustment sequence to avoid the impact of drastic weight changes on system stability.

[0033] Preferably, it also includes an A / B testing framework for verifying the performance of the new channel:

[0034] Randomly select no more than 5% of user traffic as test samples, and achieve traffic segmentation by hashing and moduloing user IDs. Distribute the test traffic to the channel to be tested and establish experimental and control groups.

[0035] The conversion data of the experimental group and the control group were collected. The confidence interval of the conversion rate was calculated based on the Beta-Binomial conjugate prior distribution. Bayesian statistical tests were performed to obtain the Bayesian statistical test results.

[0036] Based on the Bayesian statistical test results, when the difference in conversion rate between the experimental group and the control group reaches statistical significance and the relative increase exceeds 5%, the new channel that has passed the verification is automatically included in the weight allocation pool, resulting in an updated channel weight allocation pool.

[0037] Preferably, it also includes in-depth channel performance mining based on distributed asynchronous exploration:

[0038] The current channel configuration status with the optimal distribution ratio is obtained as the starting point for exploration. Parameters including channel weight, sending time period, and content template are used as exploration dimensions to construct a multi-dimensional decision tree structure containing n channel nodes and a depth of D, thereby generating a channel configuration exploration tree.

[0039] Based on the channel configuration exploration tree, k configuration mutation branches are generated in parallel, and each branch is assigned to a distributed computing node for asynchronous execution. Small-scale A / B testing is used to verify the conversion effect of each configuration scheme, and key indicator data including conversion rate and ROI are collected.

[0040] Based on the key performance indicator data, calculate the regret value for each configuration scheme. Prune inefficient branches according to the principle of minimizing regret, retaining the optimal branch for further in-depth exploration, and minimizing the most... The system converges to the globally optimal configuration within each move operation; where n is the number of channels, k is the number of parallel branches, and D is the exploration depth.

[0041] Preferably, the generation channel configuration exploration tree includes:

[0042] Set the current channel configuration state with the optimal distribution ratio as the root node, which contains a complete set of parameters including the weight allocation, sending strategy, and content template of all channels, and establish the exploration start state;

[0043] Based on the initial state of the exploration, the configuration parameters are fine-tuned to generate multiple sub-node branches, where each sub-node represents a specific variation of a certain parameter, forming a configuration variation space.

[0044] The configuration variation space is expanded to depth D according to a depth-first search strategy, with each leaf node corresponding to a specific configuration scheme, thus generating the channel configuration exploration tree.

[0045] Preferably, it also includes channel stability prediction based on a two-dimensional threshold cellular automata:

[0046] Obtain attribute information including operator type and geographic region for all SMS channels, map each channel to a two-dimensional grid structure, and establish a channel network topology model with each grid cell containing a state vector of performance score, load rate, failure probability, and neighborhood influence.

[0047] Based on the channel network topology model, a threshold conversion rule is designed. When the channel performance benchmark score is lower than 60 points and the continuous low score time exceeds 1 hour, it is judged as an early warning state. When the number of faulty channels in 8 neighborhoods is greater than or equal to 3 and the current load rate exceeds 80%, the fault probability is increased and cellular automaton state update is performed.

[0048] Based on the state update of the cellular automata, iterative evolution is performed for 100 time steps to simulate the propagation and diffusion process of channel faults, calculate the overall system stability index, and predict the stability trend of the channel network in the next 2-4 hours.

[0049] Preferably, the step of performing cellular automaton state update includes:

[0050] Obtain the state vector of each channel at the current moment, determine the state according to the performance threshold rule, the neighborhood propagation rule, and the cascading failure rule, and calculate the state probability distribution of each channel at the next moment.

[0051] Based on the state probability distribution, the state values ​​of all channels are updated in parallel, the neighborhood influence degree and fault propagation coefficient are calculated synchronously, the fault propagation process in the channel network is simulated, and the fault propagation process is obtained.

[0052] The fault propagation process is quantitatively evaluated, and stability indicators including the percentage of healthy channels and the size of the maximum connected component are calculated to form an overall system stability score.

[0053] Preferably, it also includes channel routing optimization based on a fast approximate minimum spanning tree:

[0054] Extract user characteristics, including the user group's geographical location, device type, historical behavior preferences, and active time periods, and channel characteristics, including the channel's operator type, coverage area, latency characteristics, and cost structure, to construct a high-dimensional feature space and generate a user-channel combined feature vector;

[0055] Based on the user-channel combination feature vector, the optimal connection relationship of the channel routing is constructed using the FAMST three-stage algorithm. The personalized routing network topology is established through three stages: approximate nearest neighbor graph construction, neural network component connection discovery, and iterative edge refinement.

[0056] Based on the personalized routing network topology, the shortest distance and matching degree to each channel are calculated for each sending task. The optimal channel is selected in combination with the optimal distribution ratio to achieve precise channel routing optimization at the user level.

[0057] Preferably, establishing a personalized routing network topology includes:

[0058] Based on the user-channel combination feature vector, an approximate nearest neighbor graph is constructed using Locality Sensitive Hashing, and the m=10 most similar user-channel combinations are calculated to establish an initial connection network.

[0059] The initial connection network is input into the neural network model, node features are extracted by the encoder, and the connection predictor is used to calculate the connection probability between node pairs to discover potential high-value connection relationships.

[0060] The potential high-value connections are iteratively edge-refined, the bottom 5% of edges are removed, network connectivity is checked and key connections are added, and after up to 10 iterations of optimization, the personalized routing network topology is generated.

[0061] The beneficial effects of this invention are:

[0062] 1. By using a real-time dynamic weight allocation mechanism, the limitations of traditional static configuration are overcome, enabling intelligent allocation of channel resources and dynamic optimization of SMS marketing effectiveness;

[0063] 2. Establish a multi-dimensional channel evaluation model that comprehensively considers key indicators such as reach rate, conversion rate, and complaint rate, avoiding the drawbacks of a single indicator-oriented approach and improving the comprehensiveness and accuracy of the evaluation;

[0064] 3. Integrates machine learning time series prediction capabilities, predicts channel performance trends based on LSTM neural networks, improves the foresight of decisions, and can predict channel performance changes in advance;

[0065] 4. Build an A / B testing automation framework to enable rapid verification and adaptive inclusion of new channels, reduce the cost of manual trial and error, and improve system iteration efficiency;

[0066] 5. Design anomaly detection and emergency handling mechanisms to ensure stable operation and rapid recovery of the system in the event of channel failure, thereby improving the overall reliability of the system. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1 A flowchart illustrating the SMS marketing effectiveness optimization method based on multi-channel dynamic weights provided in this embodiment of the invention;

[0069] Figure 2 This is a structural diagram of the SMS marketing performance optimization system based on multi-channel dynamic weights provided in an embodiment of the present invention. Detailed Implementation

[0070] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0071] like Figure 1 As shown in the figure, the SMS marketing effectiveness optimization method based on multi-channel dynamic weights provided in this embodiment of the invention includes the following steps:

[0072] Step S1: Obtain real-time performance data for multiple SMS channels.

[0073] In this embodiment, performance data for each SMS channel is acquired in real time through a data acquisition interface, including key performance indicators such as reach rate, conversion rate, complaint rate, and latency. Reach rate reflects the proportion of SMS messages successfully delivered to target users and is an important indicator for measuring the basic quality of the channel. Conversion rate indicates the proportion of users who perform the expected action (such as clicking a link or completing a purchase) after receiving the SMS message, directly affecting the ROI of marketing activities and the achievement of business goals. Complaint rate refers to the proportion of users who complain about received SMS messages and is an important dimension for measuring user experience and content compliance. Latency reflects the time delay from SMS sending to delivery, and is particularly crucial for scenarios with high timeliness requirements (such as verification codes and event announcements).

[0074] A distributed data acquisition architecture is adopted, acquiring real-time performance data from various SMS channels via API interfaces or database synchronization. To ensure the real-time nature and accuracy of data acquisition, an adaptive sampling strategy is implemented: under normal circumstances, a sampling period of 5 minutes is used, which is automatically increased to a high-frequency sampling of 1 minute / sample when performance fluctuations are detected, ensuring that even minor changes in channel performance can be captured. Simultaneously, the collected raw data undergoes preliminary cleaning to remove obvious outliers (such as a sudden drop in reach rate to 0%), and missing data is interpolated to ensure the data quality for subsequent analysis. Furthermore, metadata such as the timestamp and batch ID of the data acquisition are recorded for subsequent time-series analysis and data correlation.

[0075] Step S2: Standardize the real-time performance data by using the Z-score standardization method to map the reach rate, conversion rate, complaint rate, and latency index to the [0,1] interval to obtain standardized performance indexes.

[0076] Different performance indicators exhibit significant differences in their numerical range and distribution characteristics, leading to inconsistencies in measurement units when directly compared or calculated. For example, reach rate and conversion rate are typically expressed as percentages, ranging from 0% to 100%; while complaint rate is usually a smaller percentage, possibly in the range of 0% to 1%; latency indicators may be measured in seconds or milliseconds, ranging from tens of milliseconds to several seconds. These differences in magnitude and distribution mean that in comprehensive scoring calculations, indicators with larger values ​​dominate the final result, failing to objectively reflect the true importance of each indicator.

[0077] The purpose of standardization is to eliminate dimensional differences between different indicators, enabling comparison and calculation under a unified standard. Z-score standardization is a commonly used standardization method that transforms the original data into a standard normal distribution based on the mean and standard deviation. After Z-score standardization, all indicators are mapped to the same standard scale, with the numerical value representing the degree of deviation from the mean (in units of standard deviation). Subsequently, the Z-value is mapped to the [0,1] interval using the Sigmoid function, facilitating subsequent weighted calculations and scoring analysis.

[0078] After standardization, objective comparisons and weighted calculations of different performance indicators for each channel can be performed within a unified numerical range, avoiding the undue influence of differences in the original data units on the evaluation results. This step provides a standardized data foundation for subsequent comprehensive performance scoring and is a crucial preliminary step in multi-dimensional channel evaluation.

[0079] Step S3: Based on the standardized performance indicators, calculate the comprehensive performance score of each channel by weighted summation, and use the scoring formula. Generate a channel performance benchmark score.

[0080] After obtaining standardized performance metrics, it's necessary to comprehensively consider the importance of each metric to generate a comprehensive score that fully reflects the channel's performance. In real-world business scenarios, different performance metrics have varying degrees of impact on marketing effectiveness and require different weightings. For example, in campaigns primarily aimed at sales conversion, conversion rate may be more important than reach rate; while in brand promotion campaigns, low complaint rates and high reach rates may be more valued.

[0081] The overall performance score is calculated using a weighted summation method. The scoring formula is as follows: .in These represent the weighting coefficients of each indicator, satisfying... The constraints ensure the standardization of the scoring results. Since the complaint rate and delay indicators are negative indicators (the lower the value, the better), subtraction is used in the formula to ensure that the influence of each indicator on the final score is consistent.

[0082] This weighted scoring mechanism allows for dynamic adjustment of the weighting of various metrics based on different marketing scenarios and business objectives, resulting in performance scores that better reflect actual needs. For example, for verification code SMS messages with high timeliness requirements, the weighting of reach rate might be increased while the weighting of delay metrics might be decreased; for brand-sensitive marketing campaigns, the weighting of complaint rate metrics might be increased; and for sales promotion activities, the focus might be more on conversion rate metrics.

[0083] This multi-dimensional weighted scoring method overcomes the limitations of traditional single-indicator evaluation, enabling a comprehensive and objective assessment of the overall performance of each channel and providing a scientific basis for subsequent dynamic weight allocation. Furthermore, by setting performance thresholds and excellent performance standards, channel performance can be managed in a tiered manner, achieving optimal resource allocation.

[0084] Step S4: Collect multi-dimensional performance data of each channel over the past 24 hours, analyze the temporal characteristics of historical conversion data through LSTM neural network, and use a three-layer LSTM network structure to learn temporal patterns and predict the trend of conversion rate changes of each channel in the next 2-4 hours.

[0085] Channel performance is not static but dynamically changes over time. In addition to current performance scores, it's necessary to predict future performance trends to enable more proactive weighting optimization. This is especially true for conversion rate, a core business metric, as its changes directly impact the ROI and effectiveness of marketing campaigns. Predicting future conversion rate trends allows for proactive responses to potential performance fluctuations, leading to more stable marketing results.

[0086] To achieve high-precision conversion rate trend prediction, an LSTM (Long Short-Term Memory) neural network model was employed. LSTM is a special type of recurrent neural network (RNN) structure specifically designed for processing and predicting time series data. Unlike traditional feedforward neural networks, LSTM has internal states (cell states) and multiple gating mechanisms (forget gate, input gate, output gate), enabling it to effectively capture long-term dependencies and temporal patterns in the data. These characteristics make LSTM particularly suitable for analyzing and predicting time series changes in channel performance.

[0087] First, multi-dimensional performance data for each channel over the past 24 hours is collected, including conversion rate, reach rate, volume of data sent, and temporal characteristics. Then, feature engineering is performed to construct feature vectors that reflect trends across different time scales. Based on these features, a three-layer LSTM network model is trained to capture the temporal patterns and influencing factors of conversion rate changes. Finally, the trained model can predict the conversion rate trends for each channel over the next 2-4 hours, providing a forward-looking decision-making basis for subsequent weight optimization.

[0088] This machine learning-based predictive capability allows us to move beyond relying solely on current or historical performance data and instead "see" future performance changes, enabling more forward-looking dynamic optimization and further enhancing the overall effectiveness of SMS marketing.

[0089] Step S5: Based on the conversion rate trend and the channel performance benchmark score, use the gradient optimization method to solve for a weight allocation scheme with the goal of maximizing the overall ROI, and set the objective function. Calculate the optimal distribution ratio for each channel.

[0090] After collecting real-time performance data from each channel, calculating standardized performance scores, and predicting future conversion rate trends, it is necessary to determine the optimal allocation of SMS traffic across all channels to maximize overall marketing effectiveness. This is essentially a resource optimization and allocation problem, requiring consideration of the current performance status of the channels, future performance trends, and business objectives.

[0091] Mathematical optimization methods are employed to formalize the weight allocation problem into a constrained optimization problem. The objective function is set as maximizing the overall ROI, i.e., maximizing the expected overall conversion effect by optimizing the weight allocation ratio of each channel. The objective function takes the form of: ,in The weight ratio assigned to the i-th channel is represented by , the predicted conversion rate i represents the predicted conversion rate of the i-th channel in the future period, and the weight coefficient i is an adjustment parameter related to the channel cost and importance.

[0092] The optimization process needs to consider various constraints, including the requirement that the sum of all weights must be 1, that the weight of each channel must be non-negative, and an upper limit on the weight of poorly performing channels. To solve this optimization problem, the Adam optimization algorithm is adopted. This is an adaptive learning rate optimization method based on gradient descent, which can effectively handle non-convex optimization problems and provides a fast solution speed while ensuring convergence.

[0093] By solving the gradient optimization problem, the optimal weight allocation ratio for each channel is obtained. These ratios represent the resource allocation scheme that maximizes ROI under the current state and predicted trends. This dynamic weight calculation method based on mathematical optimization breaks through the limitations of traditional fixed weights or simple rules. It can automatically adjust the optimal allocation strategy according to complex and variable factors, achieving truly intelligent resource allocation.

[0094] Step S6: Based on the optimal distribution ratio, adjust the SMS queue diversion strategy in real time, redistribute the SMS messages to be sent to each channel queue, and simultaneously activate the smooth transition mechanism of weight configuration to achieve dynamic optimization of SMS marketing effectiveness.

[0095] After calculating the optimal distribution ratio, the theoretical weight configuration needs to be translated into actual SMS traffic distribution operations. Directly adjusting to the new weight configuration all at once may impact stability; therefore, a smooth transition mechanism needs to be designed to ensure stable operation during the weight adjustment process.

[0096] The core of traffic distribution adjustment is to allocate SMS messages to the sending queues of each channel in an optimal proportion. Weighted round-robin or random sampling methods are used to achieve real-time traffic distribution. For example, with a weight configuration of [A:45%, B:35%, C:20%], 45 out of every 100 SMS messages might be allocated to channel A, 35 to channel B, and 20 to channel C. To avoid excessive concentration of SMS messages on certain channels, fine-grained interleaving is typically implemented to ensure load balancing across channels.

[0097] Simultaneously, a smooth transition mechanism for weight configuration is initiated, employing a gradual adjustment strategy to avoid abrupt changes in weights. Smooth transition typically uses an exponential decay smoothing algorithm, distributing weight adjustments across multiple small steps, each step moving only a small step towards the target weight, until the final configuration is nearly complete. This gradual adjustment allows time to adapt to the new traffic distribution, avoiding instability or performance fluctuations caused by sudden traffic changes.

[0098] This dynamic adjustment and smooth transition mechanism enables continuous optimization of the overall effectiveness of SMS marketing while maintaining stable operation and reliability.

[0099] Step S2 specifically includes:

[0100] Step S21: Collect raw data on reach rate, conversion rate, complaint rate, and delay indicators for each SMS channel, calculate the mean μ and standard deviation σ of each indicator in historical data, and establish a data distribution baseline.

[0101] In this step, the raw performance metrics data for each SMS channel over the past 7 days are first collected and processed. Choosing 7 days as the historical data window is based on practical experience; it includes sufficient data samples to reflect the general distribution characteristics of the metrics while avoiding the influence of seasonal or trend changes caused by excessively long windows. For each performance metric, its historical mean μ and standard deviation σ are calculated as baseline parameters for standardization.

[0102] Taking reach rate as an example, the formula for calculating its historical average is: This mean reflects the central tendency of the indicator under normal conditions. The standard deviation is calculated as follows: The standard deviation reflects the degree of fluctuation or dispersion of the indicator.

[0103] The above calculations are performed for each performance metric for each channel, generating a parameter matrix containing the mean and standard deviation, which serves as the data distribution baseline. This baseline is used not only for subsequent standardization but also as a reference for anomaly detection. To adapt to dynamic changes in the metric distribution, a sliding window mechanism is used to periodically update these parameters, ensuring that the standardization process is based on the latest data distribution characteristics.

[0104] Simultaneously, historical quantile information for each indicator (such as the 25th, 50th, 75th, and 95th percentiles) is preserved to understand the degree of data skewness and outlier identification, providing richer statistical references for subsequent standardization mapping. This step establishes a comprehensive data distribution baseline, laying the statistical foundation for subsequent standardization processing.

[0105] Step S22: Based on the data distribution baseline, apply the Z-score normalization formula:

[0106] Normalization transformation is performed on each indicator to eliminate the influence of differences in the dimensions of different indicators, and the normalization transformation result is obtained.

[0107] In this step, the real-time acquired performance index data for each channel are normalized using the Z-score normalization formula. Z-score normalization is a linear transformation method based on the mean and standard deviation, which converts the original data points into their degree of deviation from the mean (in units of standard deviation). The Z-score formula is defined as: , where X is the original data value, μ is the historical mean of the indicator, and σ is the historical standard deviation of the indicator.

[0108] Taking reach rate as an example, if the current reach rate of a certain channel is 95%, while the historical average reach rate μ is 90% and the standard deviation reach rate σ is 3%, then the Z-value of the current reach rate of this channel is calculated as follows: This Z-score indicates that the current reach rate is 1.67 standard deviations higher than the historical average, which is considered a good performance.

[0109] For indicators such as the complaint rate, where "the lower the value, the better," the opposite approach is adopted, and the original formula is adjusted as follows: To ensure that a higher Z-value indicates better performance, consistency with other metrics is maintained. For example, if a channel's current complaint rate is 0.05%, while the historical average complaint rate μ is 0.1% and the standard deviation complaint rate σ is 0.03%, then the Z-value for the channel's current complaint rate is calculated as follows:

[0110] This indicates that the current complaint rate is 1.67 standard deviations lower than the historical average, which is considered a relatively good performance.

[0111] Z-score standardization transforms indicators with different dimensions and distribution ranges into Z-values ​​of a standard normal distribution, eliminating dimensional differences between indicators and enabling direct comparison and comprehensive calculation. This normalization transformation preserves the relative relationships and statistical characteristics of the original data while providing a unified metric, laying the foundation for subsequent performance scoring.

[0112] Step S23: Perform interval mapping on the normalization transformation result, and use the Sigmoid function to map the Z value to the standard interval [0,1] to obtain the standardized performance index.

[0113] The standardized Z-score is theoretically a variable that follows a standard normal distribution, with a range of (-∞, +∞). To facilitate subsequent weighted calculations and result interpretation, the Z-value needs to be further mapped to a bounded interval. In this step, the Sigmoid function is used to map the Z-value to the (0,1) interval. The Sigmoid function is a commonly used S-shaped curve function, defined as follows: .

[0114] The Sigmoid function has the following characteristics: when Z approaches negative infinity, f(Z) approaches 0; when Z approaches positive infinity, f(Z) approaches 1; and when Z=0, f(Z)=0.5. This means that performance similar to the historical mean will receive a standardized score of 0.5, performance better than the mean will receive a score higher than 0.5, and performance worse than the mean will receive a score lower than 0.5. Due to the non-linear nature of the Sigmoid function, extreme Z values ​​(such as ±3 and above) are compressed into a range close to 0 or 1, avoiding excessive influence of extreme values ​​on subsequent calculations.

[0115] Taking the aforementioned reach rate as an example, Z reach rate = 1.67, calculated using the Sigmoid function: This standardized score of 0.84 indicates that the channel's reach performance is better than 84% of historical data.

[0116] For metrics like complaint rate and delay rate, where "the smaller the value, the better," the Sigmoid function can be used directly for mapping after applying the adjusted Z-score formula, maintaining a consistent scoring logic (the larger the value, the better the performance).

[0117] By using the Sigmoid function to map the Z-values ​​of each indicator, they are converted into standardized performance metrics within the (0,1) interval. These metrics have a unified numerical range and interpretation method, facilitating subsequent weighted calculations and comprehensive scoring. Simultaneously, this mapping method makes the scoring results easier to understand and compare, providing a standardized indicator system for multi-dimensional evaluation of channel performance.

[0118] Step S3 specifically includes:

[0119] Step S31: Obtain business target configuration parameters and set the weight coefficients of each indicator according to the marketing scenario, including the reach rate weight. Conversion rate weight Complaint rate weighting Delay index weight satisfy Establish a weight configuration scheme.

[0120] In this step, weighting coefficients for each performance metric are set based on current business objectives and the characteristics of the marketing scenario. The weighting configuration reflects the relative importance of each metric in a specific business scenario and directly affects the channel's overall performance score. To accommodate different types of marketing campaigns, multiple weighting configuration schemes are preset, and marketers can customize them according to their specific needs.

[0121] For sales promotion marketing campaigns (such as promotional offers, limited-time discounts, etc.), set a higher conversion rate weight, configured as follows: =0.3 (reach rate) =0.5 (conversion rate) =0.15 (complaint rate) =0.05 (Delay Metric). The main goal of this type of activity is to achieve sales conversion, so the conversion rate metric has the highest weight, accounting for 50%; at the same time, to ensure basic reach quality and user experience, 30% weight is allocated to reach rate and 15% weight to complaint rate; due to the relatively low timeliness requirements, the delay metric only accounts for 5% of the weight.

[0122] For brand maintenance marketing activities (such as brand promotion, member care, etc.), the focus is more on user experience and brand image, and the configuration is as follows: =0.25 (reach rate) =0.3 (conversion rate) =0.35 (complaint rate) =0.1 (Delay Metric). These types of activities require strict control of the complaint rate to avoid negative impacts on the brand image; therefore, the complaint rate metric has the highest weight, accounting for 35%. Simultaneously, maintaining a certain level of conversion effectiveness and reach quality is crucial, with 30% weight allocated to the conversion rate and 25% to the reach rate. The delay metric has a weight of 10% to ensure timely delivery of information.

[0123] For urgent notification marketing campaigns (such as verification codes, alerts, etc.), emphasizing high reach and low latency, the configuration is as follows: =0.5 (reach rate) =0.2 (conversion rate) =0.1 (complaint rate) =0.2 (Delay Metric). The most important aspect of this type of campaign is ensuring that information is delivered to users quickly and accurately; therefore, the reach rate metric has the highest weight, accounting for 50%. The delay metric is also relatively important, accounting for 20% of the weight. Conversion rate and complaint rate are relatively less important, accounting for 20% and 10% of the weight, respectively.

[0124] The configuration management interface allows business personnel to adjust these preset schemes or create new weight configurations according to actual needs. All configuration schemes must meet the constraint that the total weight is 1, ensuring the standardization and comparability of the scoring results. Through this flexible weight configuration mechanism, channel performance scores that best meet business objectives can be generated for different marketing scenarios.

[0125] Step S32: Based on the weight configuration scheme and the standardized performance index, substitute into the scoring formula: Calculations are performed to obtain the real-time performance score for each channel.

[0126] In this step, the standardized performance indicators obtained in step S23 are combined with the weighting coefficients set in step S31, and the comprehensive performance score of each channel is calculated by weighted summation. During the calculation process, it is first ensured that all standardized indicators have been mapped to the [0,1] interval, and indicators that are "better the smaller the value" (such as complaint rate and latency indicators) have been appropriately transformed to make the direction of all indicators consistent (the larger the value, the better the performance).

[0127] The specific scoring formula is as follows:

[0128] Note that the complaint rate and delay metrics have already been processed during the standardization phase (via...). (And the Sigmoid function), here we can directly use the standardized value without needing to perform a 1-x transformation. Therefore, the actual formula simplifies to: .

[0129] To make the final scoring results more intuitive, the weighted sum is multiplied by 100 to convert it into a score range of 0-100. For example, the standardized metrics for a certain channel are: reach rate 0.85, conversion rate 0.72, complaint rate 0.93 (converted to a higher value is better), and latency index 0.88 (converted to a higher value is better). Under the weighting configuration of a sales promotion marketing campaign ( =0.3, =0.5, =0.15, =0.05), its performance score is calculated as follows: Score = (0.3×0.85 + 0.5×0.72 + 0.15×0.93 + 0.05×0.88)×100 = 79.6 points.

[0130] The above calculations are performed on all channels to obtain a set of scores reflecting the current overall performance of each channel. These scores consider multiple performance dimensions and their relative importance, providing a comprehensive and objective reflection of the overall performance of each channel in a specific business scenario. The scores are updated in real time and dynamically change with the collection of new performance data, providing the latest decision-making basis for subsequent weight optimization and traffic distribution adjustments.

[0131] Step S33: Based on the real-time performance score, set a passing threshold of 60 points and an excellent threshold of 85 points for the channel performance, and generate the channel performance benchmark score.

[0132] In this step, based on business needs and practical experience, two key performance scoring thresholds were set: a passing threshold of 60 points and an excellent threshold of 85 points. These two thresholds categorize channel performance into three levels: failing (<60 points), acceptable (60-85 points), and excellent (>85 points). The threshold settings took into account industry standards, historical data analysis, and business requirements, aiming to establish an objective channel performance grading standard.

[0133] A passing score of 60 represents the minimum acceptable level of channel performance. Channels falling below this threshold are considered poorly performing, and a series of measures will be taken to restrict their use, including: reducing their weight allocation ratio (usually not exceeding 10%), triggering performance warning notifications, and automatically activating backup channels. In extreme cases (such as multiple consecutive periods below the passing score), their weight may be reduced to 0%, completely suspending the use of the channel, and notifying operations personnel for manual intervention and channel status checks.

[0134] A score of 85 represents an excellent channel performance level. Channels exceeding this threshold are considered high-quality channels and will be given priority consideration for increased weight allocation to improve overall marketing effectiveness. For high-performing channels, their configuration parameters and usage scenarios will be recorded to create best practice cases, providing a reference for future channel selection and optimization.

[0135] Based on these two thresholds and the real-time performance score, a performance benchmark score is generated for each channel. This is a comprehensive evaluation result that includes both a score and a grade. For example, if a channel's real-time performance score is 79.6, then its performance benchmark score is "79.6 (Pass)". These benchmark scores are not only used for subsequent weight optimization calculations, but also serve as key indicators on the channel performance monitoring dashboard, visually displaying the current status of each channel.

[0136] By setting clear performance thresholds and grading standards, standardized management of channel performance is achieved, providing an objective basis for dynamic weight adjustment. It also makes it easier for operators to quickly identify problematic and high-quality channels, thereby improving overall operational efficiency.

[0137] Step S4 specifically includes:

[0138] Step S41: Based on the multi-dimensional performance data, calculate the moving average of the conversion rate over 3 hours, 6 hours, and 24 hours, construct trend features such as the first-order difference and second-order difference of the conversion rate, and form the input feature vector of the prediction model.

[0139] In this step, feature engineering is performed on the performance data of each channel collected over the past 24 hours to extract and construct feature vectors that reflect the temporal characteristics, providing high-quality input data for the LSTM prediction model. Feature engineering is crucial for the predictive performance of machine learning; good feature design can significantly improve the model's prediction accuracy.

[0140] First, the moving average for different time windows is calculated to capture trend information at different time scales. Specifically, this includes:

[0141] 3-hour moving average: Reflects short-term trend changes and is sensitive to recent fluctuations;

[0142] 6-hour moving average: reflects the medium-term trend and can smooth out the impact of short-term fluctuations;

[0143] 24-hour moving average: Reflects long-term trends and represents the average level over a complete daily cycle.

[0144] These three moving averages at different time scales together constitute a multi-scale representation of the change in conversion rate, which can help the model understand the relationship between short-term fluctuations, medium-term trends and long-term baselines.

[0145] In addition to the moving average, differential features of the conversion rate were also calculated to capture information on the rate and acceleration of change:

[0146] First-order difference: The difference between the current value and the value at the previous moment, reflecting the rate of change or the slope of the trend;

[0147] Second-order difference: The difference between the first-order difference and the second-order difference, reflecting the changing acceleration or trend curvature.

[0148] These differential features can help the model identify the dynamic characteristics of conversion rate changes, such as upward trends, downward trends, accelerating changes, or decelerating changes.

[0149] In addition, a series of auxiliary features were integrated to enrich the model's input information:

[0150] Time features: hour, day of the week, weekday / weekend markers, holiday markers, etc.;

[0151] Channel load characteristics: current transmission volume, queue length, load, etc.;

[0152] Historical cyclical pattern: historical conversion rate data for the same period;

[0153] Special event markers: marketing campaigns, maintenance, policy changes, etc.

[0154] For categorical features (such as days of the week), One-Hot encoding is used to convert them into numerical representations; for numerical features, normalization is performed to ensure consistent scale across features. Finally, a high-dimensional feature vector is generated for each channel at each time point, containing rich temporal information and contextual features, providing comprehensive input data for the LSTM model and supporting high-precision conversion rate trend prediction.

[0155] Step S42: Based on the input feature vector of the prediction model, design a three-layer LSTM network structure, wherein the input layer receives feature vectors of 24 time windows, and the hidden layer captures long-short-term dependencies through 128 LSTM units to establish a time series prediction model.

[0156] In this step, based on the feature vectors generated in step S41, a three-layer LSTM neural network model is constructed to capture complex patterns and dependencies in the conversion rate time series data. LSTM (Long Short-Term Memory) is a special recurrent neural network structure specifically designed to process sequential data. It can effectively solve the gradient vanishing problem faced by traditional RNNs and capture long-term dependencies in the data.

[0157] The designed three-layer LSTM network structure is as follows:

[0158] Input layer: Receives feature vectors from 24 consecutive time windows (each window is 1 hour). The feature vector of each time window has dimension n (containing all the features mentioned above), therefore the shape of the input tensor is [batch_size, 24, n]. The input layer feeds these features into subsequent LSTM layers for processing.

[0159] The first hidden layer, containing 128 LSTM units, is the main feature extraction layer of the network. Each LSTM unit contains a memory cell and three gating mechanisms (forget gate, input gate, and output gate), enabling it to learn when to remember or forget information. The main task of the first LSTM layer is to extract basic temporal patterns, such as short-term trends and periodic changes, from the original feature sequence. The output dimension of this layer is [batch_size, 24, 128], preserving complete temporal information.

[0160] The second hidden layer contains 64 LSTM units to further extract more advanced temporal features. Building upon the basic features extracted in the first layer, the second LSTM layer learns more complex temporal patterns, such as interactions between different features and mid-term trend changes. The output dimension of this layer is [batch_size, 24, 64].

[0161] The third hidden layer, containing 32 LSTM units, is the network's final feature fusion layer. This third LSTM layer integrates the features extracted from the first two layers, capturing the highest-level temporal patterns, such as long-term trends and the impact of special events. To accommodate multi-step prediction needs, this layer returns the output at the last moment, with dimensions [batch_size, 32].

[0162] Output layer: Contains 4 neurons, predicting conversion rates for the next 1, 2, 3, and 4 hours respectively. The output layer uses a fully connected structure, mapping the output of the third LSTM layer to the predicted values. To ensure the reasonableness of the prediction results, the output layer uses the Sigmoid activation function, constraining the predicted values ​​within the range of [0,1], which conforms to the actual meaning of conversion rate.

[0163] The network training uses mean squared error (MSE) as the loss function and the Adam optimizer for parameter updates, with an initial learning rate of 0.001. To prevent overfitting, dropout layers are added between LSTM layers with a dropout rate of 0.2. An early stopping strategy is also employed, stopping training when the validation set loss shows no improvement for five consecutive epochs. During training, a mini-batch size of 32 is used, with a maximum training epoch count of 100.

[0164] This deep LSTM network structure effectively captures various temporal patterns and dependencies in conversion rate data, achieving high-precision trend prediction. The model can not only learn obvious seasonal and periodic patterns but also adapt to sudden events and abnormal changes, providing a reliable predictive basis for subsequent weight optimization.

[0165] Step S43: Deploy the time-series prediction model as an online inference service, and combine it with the standardized performance indicators to predict the trend changes in the conversion rate of each channel in the next 2-4 hours.

[0166] In this step, the trained LSTM model is deployed as an online inference service to enable real-time prediction of future conversion rates for each channel. The model deployment employs a lightweight service architecture to ensure low latency and high availability during the prediction process, supporting weight optimization decisions every 15 minutes.

[0167] The deployment process includes the following key steps:

[0168] First, the parameters of the trained LSTM model are exported and converted into a format suitable for online inference. To improve inference efficiency, model quantization technology is used to convert 32-bit floating-point parameters into 8-bit integer representations, which significantly reduces the model size and improves computation speed while maintaining prediction accuracy.

[0169] Secondly, a prediction service cluster with auto-scaling capabilities was built, and a containerized deployment approach was adopted to ensure the elasticity and reliability of the service. The prediction service receives real-time feature data, performs model inference, and returns conversion rate predictions for future time points. To handle sudden spikes in prediction requests, a request queue and priority mechanism were implemented to ensure that prediction tasks on critical channels are processed first.

[0170] In the actual prediction process, the latest 24-hour performance data for each channel is first collected, and the same feature engineering process as in the training phase is performed to generate the input feature vectors required by the model. Then, the feature vectors are fed into the LSTM model to obtain the conversion rate prediction values ​​for the next 1-4 hours. The prediction results undergo post-processing, including reasonableness checks (ensuring that the predicted values ​​are within a reasonable range) and smoothing (avoiding drastic fluctuations in the predicted values).

[0171] It also visualizes the prediction results, allowing operations staff to intuitively understand the future trends of each channel. The visualizations include trend charts, prediction confidence intervals, historical prediction accuracy, and other information, helping operations staff assess the reliability of the predictions and formulate response strategies.

[0172] In addition to directly predicting the absolute value of the conversion rate, the algorithm also calculates the trend of conversion rate changes, classifying each channel into three categories: "upward trend," "stable trend," and "downward trend." This trend classification is more intuitive and facilitates adjustments made by the weight optimization algorithm. For example, for channels with a clear upward trend in predicted conversion rates, their weight allocation may be increased in advance; while for channels with a significant downward trend in predicted conversion rates, their weight percentage may be appropriately reduced.

[0173] By deploying the LSTM prediction model as an online service, real-time prediction of channel future performance is achieved, providing a forward-looking decision-making basis for dynamic weight optimization and significantly improving the ability to predict and respond to changes in channel performance.

[0174] Step S5 specifically includes:

[0175] Step S51: Establish the objective function for weight optimization Set constraints Construct a mathematical model for the optimization problem.

[0176] In this step, the channel weight allocation problem is formalized into a mathematical optimization model, with clearly defined optimization objectives and constraints. The objective function is set as maximizing the overall ROI, that is, maximizing the overall conversion effect by optimizing the weight allocation of each channel.

[0177] The objective function is defined as the weighted sum of the channel weights, their predicted conversion rates, and their weight coefficients:

[0178]

[0179] The formula for calculating the weighting coefficient is:

[0180] Weighting coefficient i = Base coefficient i × quality factor i × Cost Factor i

[0181] in, The predicted conversion rate represents the weight ratio assigned to the i-th channel. i The weighting coefficient represents the average conversion rate of the i-th channel over the next 2-4 hours. i This represents the adjustment parameter related to the channel characteristics in the i-th channel, and the weighting coefficient... i =Basic coefficient i × quality factor i ×Cost Factors i The quality factor i The base coefficient is calculated based on the channel performance benchmark score of the i-th channel. i The cost factor is used to reflect the fundamental importance of the i-th channel. i It is inversely proportional to the transmission cost of the i-th channel.

[0182] The constraints of the optimization model include:

[0183] 1. The sum of all channel weights is 1: ;

[0184] This ensures the reasonableness of the weight allocation, with the total weight always being 100%.

[0185] 2. The weight of each channel is non-negative: , ;

[0186] This is a constraint in a physical sense; the weight cannot be negative.

[0187] 3. Channels with performance scores below the passing grade have their weight capped: ;

[0188] To control risk, the maximum weight of poorly performing channels is limited to no more than 10%.

[0189] 4. Limitations on the magnitude of weight changes: , ;

[0190] To avoid drastic changes in weights impacting stability, the weight changes for each adjustment will not exceed 20 percentage points.

[0191] 5. Minimum weight guarantee: ;

[0192] To maintain the activity of channels and meet data collection needs, activated channels should retain a minimum weight of 5%.

[0193] These constraints ensure that the weight optimization result not only maximizes the theoretical ROI but also guarantees the feasibility and stability of practical operation. The entire optimization problem can be expressed as: It satisfies the above five constraints.

[0194] Step S52: Based on the mathematical model of the optimization problem, the Adam optimization algorithm is used to solve it. The learning rate is set to 0.01 and the maximum number of iterations is 100. The weight allocation vector that converges to the optimal solution is calculated.

[0195] In this step, the Adam (Adaptive Moment Estimation) optimization algorithm is used to solve the above optimization problem. The Adam algorithm combines the advantages of momentum method and RMSProp, and can adaptively adjust the learning rate of each parameter, making it suitable for handling non-convex optimization problems.

[0196] When implementing Adam optimization, the following specific parameters were set:

[0197] Learning rate α: 0.01, controlling the step size for parameter updates;

[0198] The exponential decay rate β1 estimated by the first moment is 0.9, which controls the degree of influence of momentum;

[0199] The exponential decay rate β2 of the second moment estimation is 0.999, which controls the smoothness of the gradient variance.

[0200] Numerical stability constant ε: 10 -8 To prevent division by zero errors;

[0201] Maximum number of iterations: 100, which controls the computational load of the algorithm;

[0202] Convergence condition: The objective function value changes by less than 10 over 5 consecutive iterations. -4 .

[0203] The optimization process starts with the current weight configuration and gradually approaches the optimal solution through iterative updates. In each iteration, the gradient of the objective function with respect to each weight parameter is calculated, and then the Adam update rule is applied to adjust the weight values. To handle constraints, a projective gradient descent method is used, projecting the weights into the feasible region after each update to ensure that all constraints are satisfied.

[0204] In practice, the non-negativity of weights and the upper and lower bound constraints of weights are first addressed by pruning operations to limit the weights of each channel within the allowable range. Then, normalization operations are used to ensure that the sum of all weights is 1. This projection method is simple and efficient, and can ensure that the constraints are met without affecting the convergence of the algorithm.

[0205] The Adam optimization algorithm can efficiently solve weight optimization problems, finding the weight allocation scheme that maximizes the overall ROI under current conditions. The algorithm typically converges to the optimal or near-optimal solution within 20-50 iterations, with a computation time on the order of 100 milliseconds, meeting the needs of real-time decision-making.

[0206] Step S53: Obtain the weight allocation vector, compare and analyze it with the current weight configuration, and output the optimal distribution ratio.

[0207] In this step, the weight allocation vector obtained by the optimization algorithm is compared and analyzed in detail with the currently used weight configuration. Weight adjustment suggestions and explanation reports are generated, and the optimal distribution ratio is finally output as the basis for the next step.

[0208] First, calculate the differences between the old and new weighting configurations, including absolute differences (percentage point changes) and relative differences (percentage changes). For channels with significant weighting changes (absolute differences exceeding 5 percentage points or relative differences exceeding 20%), they will be specially marked and the reasons for the changes will be analyzed, mainly from the following aspects:

[0209] Performance rating change: Whether the overall performance rating of the channel has changed significantly, such as being upgraded from "pass" to "excellent" or downgraded from "pass" to "fail".

[0210] Predicting conversion rate trends: Whether the future conversion rate predicted by the LSTM model will increase, remain stable, or decrease, and the strength of the trend.

[0211] Weighting coefficient adjustment: Whether the cost, quality, or other factors of the channel change, leading to an adjustment of the weighting coefficient.

[0212] Constraint impact: Whether the optimal solution is affected by constraints (such as minimum weight guarantee, maximum change range limit, etc.).

[0213] For example, the following analysis report might be generated:

[0214] "The weight of Channel A has increased from 25% to 35% (+10 percentage points, a relative increase of 40%). The main reasons are: 1) Performance score improved from 78 to 87, reaching an excellent level; 2) The predicted conversion rate for the next 3 hours shows a significant upward trend, with an expected increase of 15%; 3) The cost factor remains stable."

[0215] "The weight of Channel B decreased from 30% to 20% (-10 percentage points, a relative reduction of 33%). The main reasons are: 1) the performance score dropped from 65 to 58, below the passing grade; 2) the predicted future conversion rate shows a downward trend; and 3) due to the constraint of minimum weight guarantee, it could not be further reduced."

[0216] After completing the detailed analysis, the final optimal distribution ratio is output, which is a complete configuration including all channels and their corresponding weights. For example, for three channels A, B, and C, the optimal distribution ratio might be expressed as [A:45%, B:35%, C:20%], meaning that 45% of SMS messages are allocated to channel A, 35% to channel B, and 20% to channel C.

[0217] This optimal distribution ratio will not only serve as a direct basis for subsequent traffic adjustments, but will also be recorded in the logs for historical tracking and performance analysis. The effects of historical weight adjustments will be reviewed periodically (e.g., weekly) to evaluate the accuracy and effectiveness of the weight optimization algorithm, and to continuously improve the optimization model and parameter settings.

[0218] Through detailed comparative analysis and clear explanations of weight changes, it not only provides data-driven decision-making results but also enhances the interpretability and transparency of the decision-making process, facilitating automated decision-making that is easy for operations personnel to understand and monitor.

[0219] The smooth transition mechanism for simultaneously activating weight configuration includes:

[0220] Step S61: Obtain the optimal distribution ratio, adopt the exponential decay smoothing algorithm, set the smoothing coefficient α to 0.3, and establish the weight smoothing transition formula.

[0221] In this step, an exponential decay smoothing algorithm is used to achieve a smooth transition of weights, avoiding the impact of abrupt changes in weight allocation on stability. The smooth transition formula is defined as:

[0222] Smoothed weight = α × updated weight + (1-α) × current weight;

[0223] Here, α is the smoothing coefficient, with a value of 0.3, meaning that each adjustment only moves the target weight by 30%. In this way, the optimal weight configuration can be gradually approached, rather than being adjusted all at once.

[0224] Step S62: Based on the weight smoothing transition formula, calculate the smoothed weight = α × updated weight + (1-α) × current weight, and generate a progressive weight adjustment sequence.

[0225] In this step, the smoothed weight value for each channel is calculated according to the smooth transition formula. For multiple adjustment periods, a progressive weight adjustment sequence is generated to gradually bring the weight configuration closer to the optimal solution.

[0226] For example, assuming the current weight of channel A is 0.3 and the optimal weight is 0.45, the smoothed weight sequence is as follows:

[0227] First adjustment: 0.3 + 0.3 × (0.45 - 0.3) = 0.345;

[0228] Second adjustment: 0.345 + 0.3 × (0.45 - 0.345) = 0.3765;

[0229] Third adjustment: 0.3765 + 0.3 × (0.45 - 0.3765) = 0.39855.

[0230] Through several minor adjustments, the weight of channel A was gradually increased from 0.3 to nearly 0.45.

[0231] Step S63: Perform the weight switching operation step by step according to the progressive weight adjustment sequence to avoid the impact of drastic weight changes on system stability.

[0232] In this step, the system adjusts the sequence based on the calculated progressive weights and performs the actual weight switching operation. The specific implementation is as follows:

[0233] A weight adjustment is performed every 15 minutes, updating the configuration according to the next weight value in the sequence;

[0234] For each adjustment, the system monitors the changes in key performance indicators (such as system load, queue length, and channel response time);

[0235] If abnormal fluctuations are detected (indicator changes exceed preset thresholds), the adjustment will be paused and rolled back to the previous stable configuration;

[0236] The adjustment process is complete when the difference between the weight configuration and the target weight is less than a preset threshold (e.g., 1%).

[0237] This step-by-step switching strategy allows for a smooth transition to the new weight configuration while maintaining stable system operation.

[0238] In this embodiment, in addition to the basic process described above, the method also includes the following extended functions:

[0239] 1. A / B testing framework for verifying the performance of new channels:

[0240] Randomly select no more than 5% of user traffic as test samples, and achieve traffic segmentation by hashing and moduloing user IDs. Distribute the test traffic to the channel to be tested and establish experimental and control groups.

[0241] The conversion data of the experimental group and the control group were collected. The confidence interval of the conversion rate was calculated based on the Beta-Binomial conjugate prior distribution. Bayesian statistical tests were performed to obtain the Bayesian statistical test results.

[0242] Based on the Bayesian statistical test results, when the difference in conversion rate between the experimental group and the control group reaches statistical significance and the relative increase exceeds 5%, the new channel that has passed the verification is automatically included in the weight allocation pool, resulting in an updated channel weight allocation pool.

[0243] Specifically, the A / B testing framework is a key mechanism for identifying and validating high-performance channels, particularly suitable for evaluating the performance of newly accessed channels. Traditional methods often rely on empirical judgment or simple historical data comparisons, making it difficult to accurately quantify the statistical significance of differences in channel performance and susceptible to sample bias and random fluctuations. This system designs a rigorous and scientific A / B testing framework based on Bayesian statistical principles, enabling accurate evaluation and automated decision-making for the performance of new channels.

[0244] The framework first employs a scientific traffic segmentation method, randomly selecting no more than 5% of user traffic as a test sample. This proportion is carefully designed to provide a sufficient sample size to ensure statistical reliability without significantly impacting overall business performance, even if the test channel performs poorly. Traffic segmentation is implemented using a user ID hash modulo method, with the specific formula: user ID hash value % 100 < 5. This consistent allocation based on user ID ensures that all messages from the same user are routed to the same group (experimental group or control group), avoiding inconsistencies in user experience.

[0245] Test traffic is allocated to the channels under test to create an experimental group, while the remaining traffic uses the existing optimal channel as a control group. To ensure fairness, the system controls several key variables, including: time distribution (ensuring consistent traffic proportions between the experimental and control groups across different time periods), message type distribution (ensuring equal proportions of marketing messages, notification messages, etc., between the two groups), and user attribute distribution (ensuring similar distributions of user attributes such as geographic location and device type between the two groups). This multi-dimensional balanced allocation significantly reduces potential confounding factors and improves the reliability of test results.

[0246] During the testing process, conversion data was collected from both the experimental and control groups, including complete funnel data such as send volume, reach, click volume, and conversion volume. Data collection employed a real-time stream processing architecture to ensure that test results were promptly reflected in the decision-making process. The system not only recorded absolute values ​​but also calculated the conversion rate and relative rate of change at each stage, comprehensively evaluating channel performance.

[0247] One innovation of this system lies in its use of a Bayesian statistical method based on the Beta-Binomial conjugate prior distribution for testing and evaluation. This is more suitable for marketing scenarios than traditional frequentist hypothesis testing. The core idea of ​​the Beta-Binomial model is to treat the conversion rate as a probability distribution rather than a single-point estimate, which better expresses the uncertainty of the estimate and the incorporation of prior knowledge. In specific implementation, the system uses a Beta(α, β) distribution as the prior distribution of the conversion rate, where the α and β parameters are set based on historical data or domain knowledge. For example, for general SMS marketing, Beta(3, 97) might be used as the prior, reflecting an expected conversion rate of approximately 3%.

[0248] After collecting the experimental data, the posterior distribution is calculated using the Bayesian update rule: if the experimental group observes k conversions and the total sample size is n, then the posterior distribution is Beta(α+k, β+nk). Similarly, the posterior distribution of the control group is calculated in a similar manner. The system further calculates the probability (called the Bayesian factor) that the conversion rate of the experimental group is higher than that of the control group. This probability directly quantifies the confidence that the new channel is superior to the existing channels.

[0249] A significant advantage of Bayesian statistical tests is their ability to directly interpret the business implications of results. For example, "the probability that the new channel is better than the existing channel is 95%, with an expected conversion rate increase of 5.2%" is more intuitive and instructive for decision-making than traditional p-values. The system employs dual criteria: statistical significance (the probability that the new channel outperforms the existing channel exceeds 95%) and business significance (a relative improvement exceeding 5%). Only when both conditions are met will a new channel be considered for inclusion in the official resource pool.

[0250] For new channels that pass testing, a large amount of traffic will not be allocated hastily. Instead, a gradual weighting strategy will be adopted. The new channel will first receive a conservative initial weight (usually 5-10%), and then the weight will be gradually adjusted based on its continuous performance. The system will pay special attention to the performance stability of the new channel, including the consistency of performance across different time periods and user groups, to ensure that its advantages are generally applicable rather than accidental results in specific scenarios.

[0251] The A / B testing framework also includes automated test management features, enabling parallel testing of multiple channels simultaneously. It intelligently allocates test resources, prioritizing testing of the most promising channels. The system dynamically adjusts test durations based on initial results, prematurely ending tests on clearly excellent or poor channels to concentrate resources on boundary cases requiring more data.

[0252] This scientifically sound A / B testing framework enables the continuous discovery and validation of high-performance channels, constantly optimizing the channel resource pool and achieving long-term, stable improvements in SMS marketing effectiveness. The framework satisfies both the requirements of business agility, allowing for rapid validation of new channel performance, and the scientific nature of decision-making, avoiding risks associated with subjective judgment and experience-based decisions.

[0253] 2. Anomaly Detection and Emergency Response Mechanism:

[0254] Based on historical 30-day data, the mean μ and standard deviation σ of each performance index are calculated, and an anomaly judgment benchmark is established using the 3-sigma anomaly detection algorithm.

[0255] Real-time monitoring of performance metrics data for each channel; when the real-time metrics of a channel exceed the range of μ±3σ, it is determined to be a performance anomaly and an alarm mechanism is triggered to identify the abnormal channel.

[0256] Upon detecting the abnormal channel, immediately reduce its weight to 0% and allocate it proportionally to other channels. Simultaneously, activate the manual intervention notification mechanism to ensure the overall availability of the system.

[0257] Specifically, SMS channel performance is not always stable and may be affected by various factors such as network fluctuations, operator policy adjustments, and surges in traffic during holidays, leading to anomalies. To ensure the system can maintain stable operation in the face of various anomalies, this system has designed a comprehensive anomaly detection and emergency handling mechanism, realizing a complete closed loop from anomaly warning to automatic recovery.

[0258] The anomaly detection mechanism is based on in-depth analysis of historical data, establishing normal behavior models for various performance indicators. Using complete data from the most recent 30 days as a foundation, the system calculates the mean μ and standard deviation σ of each performance indicator (including reach rate, conversion rate, complaint rate, and latency indicators), constructing a dynamically updated anomaly judgment benchmark. The 30-day time window is a repeatedly validated optimal choice, ensuring sufficient data samples for statistical reliability while adapting to seasonal changes and long-term trends in business operations.

[0259] The 3-sigma anomaly detection algorithm is used as the basic judgment mechanism. This is a classic method based on the normal distribution assumption and is applicable to the fluctuation characteristics of most performance indicators. According to this algorithm, when the real-time value of an indicator exceeds the range of μ±3σ, it is judged as an anomaly. This standard means that under the normal distribution assumption, the theoretical probability of an anomaly event is only 0.27%, effectively balancing sensitivity and specificity.

[0260] To adapt to the characteristics of different metrics, the system has made several enhancements to the basic algorithm:

[0261] For indicators that are not normally distributed (such as complaint rates, which are usually right-skewed), the system applies Box-Cox transformation or logarithmic transformation to convert the data into an approximately normal distribution before applying the 3-sigma rule.

[0262] For metrics with obvious time patterns (such as the difference in conversion rates between weekdays and weekends), the system has established a time-segmented benchmark model to ensure that anomaly detection takes time factors into account.

[0263] Different detection sensitivities were set for indicators of varying importance. Key indicators (such as reach rate) adopted a stricter 2.5-sigma standard, while secondary indicators may adopt a more lenient 3.5-sigma standard.

[0264] In addition to static threshold detection, several advanced anomaly detection algorithms have been implemented:

[0265] Rate of change detection: Monitors the first derivative (rate of change) and second derivative (acceleration of change) of the indicator to capture sudden and drastic changes, even if the absolute value is still within the normal range.

[0266] Seasonal adjustment: Using time series decomposition techniques, the data is decomposed into trend, seasonal and residual components. Anomaly detection is applied to the residual components to eliminate the influence of normal seasonal fluctuations.

[0267] Multidimensional correlation detection: Monitor the correlation patterns between multiple indicators to discover abnormal patterns that violate historical correlations. For example, a decrease in reach is usually accompanied by an increase in latency, but if only the reach decreases while the latency is normal, it may indicate a different type of problem.

[0268] Clustering anomaly detection: Perform multi-dimensional clustering analysis on channel performance to identify outliers that deviate from normal clustering and discover complex multi-dimensional anomaly patterns.

[0269] The system monitors performance metrics for each channel in real time, with an anomaly detection engine running at a high frequency of once per minute to ensure timely problem detection. When a performance anomaly is detected, the system triggers a tiered alarm mechanism:

[0270] Level 1 Alarm (Observation Level): A single indicator slightly exceeds the threshold for a short period of time. The system records the event and increases the monitoring frequency, but no intervention measures are taken at this time.

[0271] Level 2 Alarm (Warning Level): When a single indicator continuously exceeds the threshold or multiple indicators show slight anomalies at the same time, the system sends an early warning notification to the operations team and prepares emergency resources.

[0272] Level 3 Alarm (Intervention Level): When key indicators are severely abnormal or multiple indicators are significantly abnormal at the same time, the system automatically initiates the emergency handling process and sends a high-priority notification to the technical manager.

[0273] When a channel is determined to have a serious anomaly (Level 3 alarm), the emergency handling process will be triggered immediately. The first measure is to quickly reduce the weight of the abnormal channel to 0%, completely isolating the problematic channel and preventing it from continuing to affect business operations. At the same time, the system will redistribute the traffic originally allocated to the abnormal channel to other healthy channels according to the preset emergency allocation strategy. The redistribution uses a weighted algorithm based on health and capacity to ensure that no channel is suddenly subjected to excessive load.

[0274] For example, suppose channel A (weight 30%) is detected as an anomaly and its traffic needs to be reallocated to channel B (current weight 40%) and channel C (current weight 30%). Simply allocating traffic proportionally would result in channel B receiving more traffic. The increment, C obtains The incremental increase. However, if the capacity utilization of B has reached 80% while that of C is only 50%, the system will take capacity factors into account and may allocate more traffic (such as 20%) to C and only 10% to B to prevent B from approaching its capacity limit.

[0275] Emergency response is not limited to traffic redistribution; the system will also activate a series of safeguards simultaneously:

[0276] Automatic retry mechanism: For messages that have entered the abnormal channel queue but have not yet been sent, the system will automatically reroute them to the healthy channel to ensure that the messages are not lost.

[0277] Message priority adjustment: In situations of resource scarcity, the system will increase the priority of critical business messages (such as verification codes and order notifications) to ensure that important business operations are not affected.

[0278] Traffic limiting protection: In extreme cases, if the capacity of the healthy channel cannot handle all traffic, the system will activate intelligent traffic limiting and temporarily suspend the sending of marketing messages with lower priority according to the preset business priority.

[0279] Channel recovery detection: The system will periodically send a small number of detection messages to the abnormal channel (usually no more than 1% of the total traffic) to monitor its recovery. Once the system detects that the performance has recovered to the normal level, it will start a gradual weight recovery process.

[0280] Human intervention notification mechanisms are an important supplement to emergency response, ensuring that human experts provide oversight and intervention in addition to automated measures. The system sends detailed anomaly reports through multiple channels (email, SMS, instant messaging, telephone), including anomaly indicators, severity of the anomaly, possible cause analysis, and measures already taken by the system. The reports employ a layered design, providing both summary information understandable to management and detailed diagnostic data required by technical personnel.

[0281] A complete lifecycle management system for abnormal events has also been established, recording the entire process from discovery to resolution, including detection time, alarm level, scope of impact, handling measures, recovery time, and root cause analysis. These records are not only used for post-event auditing but also serve as training data for machine learning models, continuously improving the system's anomaly detection accuracy and processing efficiency.

[0282] Through this comprehensive anomaly detection and emergency response mechanism, the system maintains a high degree of resilience and flexibility in the face of various anomalies, minimizing the impact of abnormal events on business operations and ensuring the overall availability and stability of the SMS marketing system. Even under the most unfavorable conditions, the system can quickly adjust resource allocation, prioritizing critical business operations and achieving a smooth transition in business continuity.

[0283] 3. In-depth channel performance analysis based on distributed asynchronous exploration:

[0284] The current channel configuration status with the optimal distribution ratio is obtained as the starting point for exploration. Parameters including channel weight, sending time period, and content template are used as exploration dimensions to construct a multi-dimensional decision tree structure containing n channel nodes and a depth of D, thereby generating a channel configuration exploration tree.

[0285] Based on the channel configuration exploration tree, k configuration mutation branches are generated in parallel, and each branch is assigned to a distributed computing node for asynchronous execution. Small-scale A / B testing is used to verify the conversion effect of each configuration scheme, and key indicator data including conversion rate and ROI are collected.

[0286] Based on the key performance indicator data, calculate the regret value for each configuration scheme. Prune inefficient branches according to the principle of minimizing regret, retaining the optimal branch for further in-depth exploration, and minimizing the most... The system converges to the globally optimal configuration within each move operation.

[0287] Specifically, traditional channel weight optimization is usually limited to fine-tuning existing configurations, making it difficult to discover potential performance breakthroughs and innovative combinations. To overcome this limitation, this system designs a channel performance deep mining mechanism based on distributed asynchronous exploration. It adopts a combination of heuristic search and parallel computing to explore the optimal configuration in a broader parameter space and uncover the potential upper limit of channel performance.

[0288] This mechanism first obtains the current channel configuration status as the starting point for exploration, and then expands the exploration dimensions to multiple key parameters, far exceeding the traditional single weight adjustment:

[0289] Channel weight: The traffic allocation ratio of each channel, which is the most basic optimization dimension.

[0290] Sending time period: The sending priority and weight are adjusted for different time periods, taking into account the time differences in user activity and channel performance.

[0291] Content templates: The format, length, personalization, and presentation of messages have a significant impact on conversion rates.

[0292] User segmentation: Targeted messaging strategies based on user characteristics (such as activity level, purchasing power, and geographic location).

[0293] Sending frequency: Message frequency control strategies for different user groups.

[0294] Channel combination: a multi-channel collaboration strategy for specific business scenarios.

[0295] A multi-dimensional decision tree structure, called the channel configuration exploration tree, is constructed, containing n channel nodes and a depth D. Each node represents a complete set of configuration parameters, and the connections between nodes represent parameter adjustment paths. The root node of the tree is the current configuration, and each level represents an optimization dimension. The maximum depth D of the tree is typically set to 6-8 to ensure both breadth and depth of exploration.

[0296] Based on this exploration tree, k configuration mutation branches (typically k=5-10) are generated in parallel, each branch representing a possible optimization direction. Mutation generation employs a combination of strategies:

[0297] Gradient-guided mutation: Based on historical data, the gradient of the parameters with respect to the objective function is calculated, and mutations are generated along the gradient direction.

[0298] Random perturbation mutation: Add random perturbations to the current configuration to explore the local space.

[0299] Crossover mutation: Combining multiple historically well-configured parameters to obtain potential superior characteristics.

[0300] Genetic Algorithm Mutation: Borrowing from the mutation operation of genetic algorithms, a moderate degree of randomness is introduced.

[0301] Expert rule variation: Targeted variation of rules generated based on domain knowledge pre-sets.

[0302] These mutation branches are distributed across different nodes in the distributed computing cluster for asynchronous execution and evaluation. The system employs containerization technology and work queue management to ensure efficient utilization of computing resources and reliable task execution. Each computing node independently evaluates its assigned configuration scheme without waiting for other nodes, significantly improving exploration efficiency.

[0303] To evaluate the effectiveness of each configuration option, a small-scale A / B testing method was employed, allocating 1-2% of traffic to each option for experimentation. This small-scale testing provides statistically significant results while minimizing experimental risks. Key performance indicators (KPIs) data from multiple dimensions, including conversion rate, ROI, and user feedback, were collected to comprehensively evaluate the effectiveness of the configuration options.

[0304] During the exploration process, a multi-armed gambling machine strategy combining Thompson sampling and the upper confidence bound (UCB) algorithm was adopted to dynamically adjust the resource allocation of each branch. Branches with better performance received more test traffic and computing resources, while poorly performing branches were eliminated early, achieving optimal resource allocation.

[0305] For each configuration, its regret value (the difference between it and the currently known optimal solution) is calculated, and branch pruning is performed based on the principle of minimizing regret. This method ensures the efficiency of exploration and avoids over-investment in obviously suboptimal solutions. Theoretically, the system can achieve at most 2n+O(k) iterations. 2 2 KD The system converges to the globally optimal configuration within 10 moves, where n is the number of channels, k is the number of parallel branches, and D is the exploration depth.

[0306] The exploration process is not a one-off event, but a continuous, ongoing mechanism. The system will launch a new round of in-depth exploration periodically (e.g., weekly) to adapt to changes in the business environment and user behavior. Each round of exploration retains the best results from the previous round as a starting point, enabling the accumulation and transfer of knowledge.

[0307] To balance exploration and utilization, the system adopts an ε-decay strategy: exploration accounts for a larger proportion in the early stages (e.g., 20% of traffic is used for exploration), and as time goes on and performance improves, the proportion of exploration gradually decreases (it may eventually drop to 5%), so that more resources are used to utilize known excellent configurations.

[0308] A key innovation of this mechanism is its asynchronous and parallel experimental evaluation system, which overcomes the efficiency bottleneck of traditional sequential experiments. In traditional methods, only one configuration change can be tested at a time, and sufficient data must be accumulated before a judgment can be made; while this system can evaluate multiple configuration schemes simultaneously, significantly accelerating the optimization iteration speed.

[0309] It also implemented automatic analysis and knowledge extraction functions for experimental results, summarizing universal optimization rules and patterns from successful configuration schemes, such as "the conversion rate of personalized content sent to young users between 10 PM and 2 AM is significantly higher than the average level." These rules were added to the knowledge base to guide future optimization directions.

[0310] Through this deep mining mechanism based on distributed asynchronous exploration, the system can break through the limitations of local optima, seek innovative configuration combinations in a broader parameter space, and uncover the potential upper limit of channel performance. Practice has shown that this method can bring an additional 10-20% performance improvement compared to traditional single-dimensional optimization, especially in complex and ever-changing business environments.

[0311] The generation of the channel configuration exploration tree includes:

[0312] Set the current channel configuration state with the optimal distribution ratio as the root node, which contains a complete set of parameters including the weight allocation, sending strategy, and content template of all channels, and establish the exploration start state;

[0313] Based on the initial state of the exploration, the configuration parameters are fine-tuned to generate multiple sub-node branches, where each sub-node represents a specific variation of a certain parameter, forming a configuration variation space.

[0314] The configuration variation space is expanded to depth D according to a depth-first search strategy, with each leaf node corresponding to a specific configuration scheme, thus generating the channel configuration exploration tree.

[0315] Specifically, the current channel configuration state with the optimal distribution ratio is set as the root node, containing a complete set of parameters for all channels' weight allocation, sending strategies, and content templates, establishing the initial state for exploration. This step first uses the current optimal configuration of the system as the starting point for exploration, ensuring that the exploration process is based on verified effectiveness. The root node contains complete configuration information, not only limited to channel weight allocation, but also including sending priority strategies for each time period, targeting rules for various user groups, and content template parameters for different scenarios. The system comprehensively records these parameters, including their value ranges, sensitivities, and interdependencies, providing constraints and guidance for subsequent mutation exploration.

[0316] Based on the initial state of the exploration, the system fine-tunes various configuration parameters, generating multiple child node branches. Each child node represents a specific variation of a certain parameter, forming a configuration variation space. In this stage, the system employs various variation strategies to generate direct child nodes of the root node. For continuous parameters (such as channel weights), the system generates three variations: increasing, decreasing, and remaining unchanged. For discrete parameters (such as content template types), it tries all possible alternative values. The variation magnitude is dynamically adjusted according to parameter sensitivity; important parameters use smaller variation magnitudes, while secondary parameters may have larger adjustment space. The system also considers the coupling relationships between parameters to ensure that the mutated configuration still meets business constraints, such as a total weight sum of 100% and a minimum guarantee for critical channels. Through this targeted parameter fine-tuning, the system can fully explore potential optimization points in the parameter space while maintaining configuration rationality.

[0317] The configuration mutation space is expanded to depth D using a depth-first search strategy, with each leaf node corresponding to a specific configuration scheme, generating the channel configuration exploration tree. The system employs a depth-first search (DFS) strategy, expanding the exploration tree layer by layer until a preset maximum depth D is reached. During expansion, the system applies mutation rules similar to those for the root node to each intermediate node, generating child nodes for the next layer. As the tree depth increases, the accumulated mutation effects produce schemes that differ significantly from the initial configuration, allowing the exploration to cover a wider configuration space. To control the tree size and computational complexity, the system implements an intelligent pruning mechanism, prematurely terminating the expansion of obviously suboptimal branches. Pruning is based on factors including the performance of historically similar configurations, domain knowledge rules, and preliminary evaluation results. The final exploration tree typically contains hundreds to thousands of leaf nodes, each representing a complete candidate configuration scheme awaiting further evaluation and verification by the system. This structured exploration method ensures both breadth and depth of the search while improving exploration efficiency through heuristic rules, enabling the discovery of innovative optimized configurations within reasonable computational resource constraints.

[0318] 4. Channel stability prediction based on two-dimensional threshold cellular automata:

[0319] Obtain attribute information including operator type and geographic region for all SMS channels, map each channel to a two-dimensional grid structure, and establish a channel network topology model with each grid cell containing a state vector of performance score, load rate, failure probability, and neighborhood influence.

[0320] Based on the channel network topology model, a threshold conversion rule is designed. When the channel performance benchmark score is lower than 60 points and the continuous low score time exceeds 1 hour, it is judged as an early warning state. When the number of faulty channels in 8 neighborhoods is greater than or equal to 3 and the current load rate exceeds 80%, the fault probability is increased and cellular automaton state update is performed.

[0321] Based on the state update of the cellular automata, iterative evolution is performed for 100 time steps to simulate the propagation and diffusion process of channel faults, calculate the overall system stability index, and predict the stability trend of the channel network in the next 2-4 hours.

[0322] Specifically, the stability of SMS channels not only affects the performance of individual channels but also involves the systemic risk of the entire channel network. Channels have complex interdependencies and influences; a failure in one channel can propagate through various mechanisms and affect other channels, ultimately leading to systemic risk. Traditional independent monitoring methods struggle to predict such network effects. Therefore, this system innovatively introduces a channel stability prediction mechanism based on a two-dimensional threshold cellular automata, capable of simulating the dynamic process of fault propagation and providing early warnings of system risks.

[0323] Cellular automata are discrete dynamic system models composed of simple rules but complex behaviors, widely used in the simulation of complex systems. This system applies them to channel stability prediction, constructing a two-dimensional network model that reflects the relationships between channels.

[0324] First, obtain the attribute information of all SMS channels, including operator type (such as China Mobile, China Unicom, and China Telecom), geographical region (such as East China, South China, and North China), service provider, and capacity level. These attributes form the basis for constructing the channel network topology and determine the "proximity" and potential influence relationships between channels.

[0325] Based on these attributes, the system maps each channel to a two-dimensional grid structure, forming a channel network topology model. The mapping process employs a multidimensional scaling (MDS) algorithm, converting the attribute similarity between channels into distances in two-dimensional space, ensuring that channels with similar attributes are also located close together in the grid. For example, channels from the same operator and in the same region will be mapped to adjacent locations, reflecting their potential high correlation.

[0326] In this two-dimensional mesh structure, each mesh cell (representing a channel) contains a state vector that describes the key characteristics of its current state:

[0327] Performance score: Reflects the current overall performance level of the channel and is calculated by the aforementioned scoring model.

[0328] Load factor: The ratio of the current channel's load level to its maximum capacity.

[0329] Failure probability: The estimated probability of a channel failure, calculated based on historical failure data and the current state.

[0330] Neighborhood impact: The potential impact of a channel failure on surrounding channels, which is related to the channel's importance and connectivity.

[0331] Threshold-based state transition rules were designed for cellular automata to simulate the dynamic evolution of channel states over time and with changing environment. The core transition rules include:

[0332] Warning status determination: When the channel performance benchmark score is below 60 points and the low score lasts for more than 1 hour, the channel enters the warning status and the probability of failure begins to rise.

[0333] Fault propagation rule: When the number of faulty channels in the 8-neighborhood (Moore's neighborhood) is greater than or equal to 3 and the current load rate exceeds 80%, the failure probability of the central channel increases significantly. This reflects the cascading failure effect under high load conditions.

[0334] Recovery rule: When the fault pressure in the neighborhood of the channel decreases and the system takes intervention measures (such as reducing the load), the probability of failure begins to decrease, and the channel has a chance to return to normal.

[0335] Threshold adjustment rule: As time goes by, the state threshold of the channel will be dynamically adjusted according to the overall system status to simulate the adaptive behavior of the system.

[0336] With the state vector and transition rules, the system executes the state update process of a cellular automaton. At each time step, the states of all channels are updated simultaneously according to the transition rules and the current neighborhood state. This parallel update mechanism can capture emergent behavior and nonlinear dynamics in complex systems.

[0337] Typically, the system iterates for 100 time steps (each step representing approximately 2-3 minutes in a real system) to simulate the propagation of a channel fault. During this evolution, the system records the changes in key indicators, including:

[0338] Faulty Channel Ratio: The percentage of channels that are in a faulty state out of the total number of channels.

[0339] System connectivity: A graph theory connectivity metric for a network, reflecting the overall health of the channel network.

[0340] Critical channel status: The status changes of core channels that are particularly important to the system.

[0341] Clustering coefficient: reflects the degree of clustering of channel faults. High clustering means that faults are concentrated in a specific area.

[0342] Based on these evolutionary trajectories, the system calculates a series of stability indices to assess the overall stability trend of the channel network over the next 2-4 hours:

[0343] System Vulnerability Index: An overall risk measure calculated based on the probability distribution of channel failures and network structure.

[0344] Cascading failure risk: The estimated probability of a large-scale cascading failure occurring in the system.

[0345] Recovery resilience index: An estimate of a system's ability to recover from partial failures.

[0346] Stable range prediction: The expected duration for which the system will remain in a stable state.

[0347] These indicators not only provide quantitative predictions of future stability but also identify vulnerabilities and potential sources of risk in the system. For example, they might provide a warning: "Within the next 3 hours, there is a 20% risk of cascading failures in the telecommunications channels in the East China region. It is recommended to reduce the load and activate the backup channels in advance."

[0348] The prediction results are directly fed back to the weight optimization module, affecting subsequent weight allocation decisions. When the system predicts an increase in instability in a certain region or type of channel, it will proactively adjust the weights of the relevant channels, shifting traffic to more stable regions and proactively reducing risk.

[0349] To improve prediction accuracy, an adaptive learning mechanism is employed, continuously adjusting the parameters and rules of the cellular automata based on observed fault propagation patterns. This includes calibrating state transition thresholds, adjusting neighborhood influence weights, and introducing new transition rules. Through this closed-loop learning, the system can progressively improve the simulation accuracy of the dynamic characteristics of specific channel networks.

[0350] Another innovation of this prediction mechanism is its integration with external data sources. It automatically acquires and integrates various external factors that may affect channel stability, such as:

[0351] Major Events Calendar: Events that may cause a surge in traffic, such as holidays and promotional activities.

[0352] Carrier maintenance plan: Network maintenance or upgrade activities that have been announced.

[0353] Abnormal weather data: Warnings of severe weather that may affect communication infrastructure.

[0354] Internet health status: Global or regional indicators of network congestion or anomalies.

[0355] These external data serve as adjustment factors for cellular automata, influencing the probability and speed of state transitions and making the prediction model more closely reflect reality.

[0356] This channel stability prediction mechanism based on a two-dimensional threshold cellular automata enables the simulation and prediction of macroscopic network stability changes from a microscopic channel state perspective, achieving a shift from passive response to proactive prevention. This capability is invaluable for ensuring the stable operation of systems in complex and ever-changing environments, especially during peak periods and special timeframes. Predictive risk management can effectively prevent potential systemic failures and ensure business continuity.

[0357] The execution of the cellular automaton state update includes:

[0358] Obtain the state vector of each channel at the current moment, determine the state according to the performance threshold rule, the neighborhood propagation rule, and the cascading failure rule, and calculate the state probability distribution of each channel at the next moment.

[0359] Based on the state probability distribution, the state values ​​of all channels are updated in parallel, the neighborhood influence degree and fault propagation coefficient are calculated synchronously, the fault propagation process in the channel network is simulated, and the fault propagation process is obtained.

[0360] The fault propagation process is quantitatively evaluated, and stability indicators including the percentage of healthy channels and the size of the maximum connected component are calculated to form an overall system stability score.

[0361] Specifically, the system obtains the state vector of each channel at the current moment, and determines the state according to the performance threshold rule, neighborhood propagation rule, and cascading failure rule, calculating the state probability distribution of each channel at the next moment. The system first extracts the latest state vector of each channel from the monitoring database, containing four key dimensions: performance score, load rate, failure probability, and neighborhood influence. For each channel, three core rules are applied to determine state evolution: the performance threshold rule adjusts the failure probability based on the comparison between the channel's current performance score and a preset threshold (usually 60 points); the longer the score remains below the threshold, the faster the failure probability increases. The neighborhood propagation rule examines the state of the channel's eight neighboring channels (Moore's neighborhood); when the number of faulty channels in the neighborhood reaches a specific threshold (usually three) and the current channel's load rate exceeds 80%, the failure probability of that channel is significantly increased. The cascading failure rule simulates a chain reaction under high load conditions; when a channel enters a fault state, the traffic it carries is redistributed to other channels, increasing the load rate and failure risk of these channels. The system integrates the influence of these three types of rules to calculate the probability distribution of each channel's possible state (healthy, warning, minor fault, or severe fault) at the next time step. Based on this state probability distribution, the system updates the state values ​​of all channels in parallel, simultaneously calculates the neighborhood influence degree and fault propagation coefficient, and simulates the fault propagation process in the channel network to obtain the fault propagation process. The system uses the Monte Carlo method to simultaneously determine the specific state of all channels at the next time step based on the calculated state probability distribution. This parallel update mechanism can accurately simulate the complex dynamics of multiple channels evolving simultaneously in the real world. After the state update is completed, the system recalculates the neighborhood influence degree of each channel. This indicator reflects the potential impact intensity of a channel fault on surrounding channels and is positively correlated with the channel's importance (carrying capacity), connectivity (degree of association with other channels), and current state (fault severity). Simultaneously, the system also calculates the fault propagation coefficient to measure the speed and range of fault propagation in the network. This coefficient is affected by factors such as network topology, channel similarity, and load distribution. Through iterative updates across multiple time steps, the system generates a complete spatiotemporal evolution sequence of fault propagation, intuitively demonstrating how faults propagate, spread, accumulate, or dissipate within the channel network, thus forming a comprehensive simulation of the system's dynamic stability. The fault propagation process is quantitatively evaluated, calculating stability indicators including the percentage of healthy channels and the size of the largest connected component, resulting in an overall system stability score. The simulated fault propagation process is quantitatively analyzed from multiple dimensions, extracting key stability indicators. The percentage of healthy channels reflects the overall health of the system; it is calculated by dividing the number of healthy channels by the total number of channels. The trend of this indicator over time reveals whether the system is moving towards stability or instability.The maximum connected component size is calculated based on graph theory analysis, treating the channel network as a graph structure with healthy channels as nodes and the connections between channels as edges. This metric reflects the system's functional integrity and service continuity. In addition, the system calculates auxiliary metrics such as fault clustering coefficient (measuring the concentration of faults), critical channel stability (the stability of core channels), and resilience (the system's ability to recover from localized faults). Finally, the system generates an overall stability score from 0 to 100 by weighted fusion of these metrics. Based on historical data and expert experience, a warning threshold is set. When the stability score falls below a specific threshold (typically 65), a corresponding warning mechanism is triggered, guiding the operations team to take early intervention measures to prevent potential systemic risks.

[0362] 5. Channel routing optimization based on fast approximate minimum spanning tree:

[0363] Extract user characteristics, including the user group's geographical location, device type, historical behavior preferences, and active time periods, and channel characteristics, including the channel's operator type, coverage area, latency characteristics, and cost structure, to construct a high-dimensional feature space and generate a user-channel combined feature vector;

[0364] Based on the user-channel combination feature vector, the optimal connection relationship of the channel routing is constructed using the FAMST three-stage algorithm. The personalized routing network topology is established through three stages: approximate nearest neighbor graph construction, neural network component connection discovery, and iterative edge refinement.

[0365] Based on the personalized routing network topology, the shortest distance and matching degree to each channel are calculated for each sending task. The optimal channel is selected in combination with the optimal distribution ratio to achieve precise channel routing optimization at the user level.

[0366] Specifically, traditional SMS channel selection typically employs weighted random allocation or simple round-robin, neglecting the matching degree between user characteristics and channel features, resulting in poor resource allocation efficiency. This system innovatively introduces a channel routing optimization mechanism based on Fast Approximate Minimum Spanning Tree (FAMST), achieving precise channel matching at the user level and significantly improving the targeting and effectiveness of SMS sending.

[0367] This mechanism first constructs a high-dimensional feature space to comprehensively capture the multi-dimensional features of users and channels. User features include:

[0368] Geographic location: The province, city, and region where the user is located, which directly affects the network connection quality with different carrier channels.

[0369] Device type: The brand, model, and operating system version of the user's mobile phone. These factors affect the display effect and interactive experience of messages.

[0370] Historical behavioral preferences: users' historical open rate, click-through rate, conversion rate, and response patterns to different types of messages.

[0371] Active time periods: Typical active times for users, such as morning, noon, evening or late at night, reflecting users' usage habits.

[0372] User value: A comprehensive user value score calculated based on dimensions such as spending power, activity level, and loyalty.

[0373] Sensitivity: How sensitive a user is to the frequency of messages, past unsubscription or complaint behavior, etc.

[0374] Channel characteristics include:

[0375] Operator type: The telecommunications operator (such as China Mobile, China Unicom, China Telecom) or third-party service provider to which the channel belongs.

[0376] Coverage area: The geographical coverage of the channel and its performance in each region.

[0377] Latency characteristics: The latency performance of the channel under different time periods and different loads.

[0378] Content compatibility: The channel's ability to support different message formats, lengths, and rich media content.

[0379] Time-of-day performance: The performance fluctuation pattern of the channel at different times of the day.

[0380] Cost structure: Billing methods for channels and cost variations under different conditions.

[0381] Based on these characteristics, the system generates user-channel combination feature vectors to evaluate the potential matching degree of each user-channel combination. These feature vectors include cross-combinations of the original features, weight adjustments, and nonlinear transformations, forming a high-dimensional matching space.

[0382] In this feature space, the system employs the FAMST (Fast Approximate Minimum Spanning Tree) three-stage algorithm to construct the optimal connectivity for channel routing. The FAMST algorithm is an efficient approximation of the traditional minimum spanning tree algorithm, particularly suitable for handling connectivity optimization problems with large-scale datasets. Its three key stages include:

[0383] Approximate Nearest Neighbor Graph Construction: First, the Locality Sensitive Hashing (LSH) algorithm is used to quickly construct an approximate nearest neighbor graph in the feature space. LSH achieves an approximate nearest neighbor search with an O(n log n) time complexity by mapping high-dimensional data to low-dimensional signatures, greatly improving processing speed. In this step, the k most similar channels (usually k=3-5) in the feature space are found for each user, forming an initial candidate set.

[0384] Neural Network Component Connection Discovery: A trained deep neural network model is used to evaluate the quality scores of candidate user-channel connections. This model takes a user-channel combined feature vector as input and outputs a matching score between 0 and 1, reflecting the expected performance of the connection. The neural network model is trained on historical transmitted data and is able to capture complex nonlinear relationships and hidden patterns. Based on these scores, the optimal connections are retained, forming a directed weight graph.

[0385] Iterative edge refinement: The system continuously refines and optimizes connections through an iterative process. In each iteration, the system removes the edge with the lowest weight and attempts to add new potential connections, evaluating whether this improves the overall performance. This process is similar to Kruskal's algorithm in building a minimum spanning tree, but it employs heuristic rules and an early stopping strategy, significantly improving efficiency. Iteration continues until the convergence condition is met or the maximum number of iterations is reached.

[0386] Through these three stages, an optimized personalized routing network topology is ultimately constructed, determining the best channel selection strategy for each user and each message type. This topology considers both local user-channel matching and global load balancing and cost-effectiveness.

[0387] In actual channel routing decisions, the system comprehensively considers three key factors:

[0388] Matching degree: The similarity score of user-channel combinations in the feature space, reflecting the degree of fit of the channel to a specific user.

[0389] Shortest distance: In a routing network topology, the path distance from sending a task to each channel reflects the efficiency of routing.

[0390] Weight constraint: It is necessary to meet the globally optimal distribution ratio calculated in step S5 to ensure that the overall traffic allocation conforms to the optimization result.

[0391] A weighted multi-objective optimization method is employed to balance these three factors and select the optimal channel for each transmission task. The specific decision function is as follows:

[0392] ;

[0393] The weight satisfaction measure is the weight allocation that determines whether selecting a channel helps achieve the goal. For example, if a channel's current actual usage rate is lower than its target weight, then selecting that channel has a higher weight satisfaction.

[0394] To improve system efficiency, routing decisions employ multi-level caching and batch processing strategies. The system caches routing results for similar users and applies the same routing strategy to users with identical characteristics in batches, significantly reducing computational overhead. Simultaneously, an incremental update mechanism is implemented, recalculating the routing strategy only when user characteristics or channel performance change significantly.

[0395] Another innovation of this mechanism is its adaptive learning capability. It continuously collects performance data from each transmission, including reach status, user response, and final conversion, constantly updating and refining the user-channel matching model. This closed-loop learning allows the routing strategy to adapt to changes in user behavior and fluctuations in channel performance, achieving continuous optimization.

[0396] It also features real-time response capabilities, enabling dynamic adjustments to routing decisions based on current system load and channel status. For example, if a sudden increase in real-time latency is detected on a channel, the system will temporarily lower the matching score of that channel and redirect traffic to other stable channels until the abnormal situation is resolved.

[0397] This channel routing optimization mechanism, based on a fast approximate minimum spanning tree, achieves a leap from coarse-grained channel-level allocation to fine-grained user-level precise matching, significantly improving the targeting and effectiveness of SMS marketing. Practice has shown that compared to traditional random allocation, this precise matching mechanism can increase conversion rates by 15-25% while reducing complaint rates by 10-15%, achieving a dual improvement in marketing effectiveness and user experience.

[0398] The establishment of a personalized routing network topology includes:

[0399] Based on the user-channel combination feature vector, an approximate nearest neighbor graph is constructed using Locality Sensitive Hashing, and the m=10 most similar user-channel combinations are calculated to establish an initial connection network.

[0400] The initial connection network is input into the neural network model, node features are extracted by the encoder, and the connection predictor is used to calculate the connection probability between node pairs to discover potential high-value connection relationships.

[0401] The potential high-value connections are iteratively edge-refined, the bottom 5% of edges are removed, network connectivity is checked and key connections are added, and after up to 10 iterations of optimization, the personalized routing network topology is generated.

[0402] Specifically, based on the user-channel combination feature vector, an approximate nearest neighbor graph is constructed using Locality Sensitive Hashing (LSH), and the m=10 most similar user-channel combinations are calculated to establish an initial connection network. This step first maps user-channel combinations in the high-dimensional feature space to low-dimensional hash signatures, quickly identifying potential nearest neighbor relationships by comparing the similarity of the hash signatures. The system employs a multi-hash table strategy, using different families of hash functions to generate multiple hash tables, improving the accuracy and recall of nearest neighbor search. For each user-channel combination, the system retrieves the m=10 most similar combinations in the hash space, constructing preliminary connection relationships. This LSH-based approximate search reduces computational complexity from traditional... This significantly improves the ability to process large-scale data. The initial connection network forms a sparse graph structure, laying the foundation for subsequent fine-grained optimization.

[0403] The initial connection network is input into a neural network model. An encoder extracts node features, and a connection predictor calculates the connection probabilities between node pairs to identify potential high-value connections. In this stage, the system deploys a graph neural network model containing two core components: a feature encoder and a connection predictor. The feature encoder employs a multi-layer graph convolutional network (GCN) structure, capable of capturing local topological information and feature representations of nodes, converting the original feature vectors into embedded representations containing structural information. The connection predictor calculates connection probabilities based on the embedded representations of node pairs, using a multi-layer perceptron structure, and outputs a connection strength score between 0 and 1. This model undergoes supervised learning using historical data, enabling it to identify potential high-value connections missed in the initial connection network and filter out low-quality connections that LSH might introduce. The system calculates predicted scores for all possible connections, retaining connections with scores exceeding a threshold (typically set to 0.7) to form an enhanced connection network.

[0404] The system iteratively refines the edges of potential high-value connections, removing the bottom 5% of edges, checking network connectivity, and adding key connections. This process, involving up to 10 iterations, generates the personalized routing network topology. Iterative edge refinement is a fine-tuning process where the system performs three key operations in each iteration: First, it removes the bottom 5% of edges in the current network, which typically represent inefficient or redundant connections; second, it checks network connectivity after edge removal to ensure no nodes are isolated, adding key connections as needed to maintain the network's basic structure; and finally, it re-evaluates the network's overall performance metrics, including average path length, clustering coefficient, and load balancing. Optimization progress is monitored using an objective function, and optimization stops when the performance improvement after two consecutive iterations falls below a preset threshold (0.5%) or the maximum number of iterations (10). The resulting personalized routing network topology retains high-value connections while possessing good structural characteristics, supporting efficient and accurate message routing decisions. This topology is updated periodically (typically every 24 hours) to adapt to dynamic changes in user behavior and channel performance.

[0405] like Figure 2 As shown, this embodiment of the invention also provides a system for optimizing SMS marketing effectiveness based on multi-channel dynamic weights, including:

[0406] The data acquisition module 701 is used to acquire real-time performance data from multiple SMS channels;

[0407] The standardization processing module 702 is used to standardize the real-time performance data. The Z-score standardization method is used to map the reach rate, conversion rate, complaint rate and delay index to the [0,1] interval to obtain standardized performance indexes.

[0408] The scoring calculation module 703 is used to calculate the comprehensive performance score of each channel based on the standardized performance indicators by weighted summation, using the following scoring formula: Generate channel performance benchmark scores;

[0409] The prediction module 704 is used to collect multi-dimensional performance data of each channel in the past 24 hours, analyze the temporal characteristics of historical conversion data through LSTM neural network, and use a three-layer LSTM network structure to learn temporal patterns and predict the trend of conversion rate changes of each channel in the next 2-4 hours.

[0410] The weight optimization module 705 is used to solve a weight allocation scheme with the goal of maximizing the overall ROI based on the conversion rate trend change and the channel performance benchmark score, using a gradient optimization method, and setting an objective function. Calculate the optimal distribution ratio for each channel;

[0411] The traffic distribution adjustment module 706 is used to adjust the traffic distribution strategy of the SMS queue in real time according to the optimal distribution ratio, redistribute the SMS messages to be sent to each channel queue, and at the same time activate the smooth transition mechanism of weight configuration to achieve dynamic optimization of SMS marketing effect.

[0412] The functions of each module in this embodiment correspond to the aforementioned method steps, achieving SMS marketing effectiveness optimization based on multi-channel dynamic weights through collaborative work. This system can automatically complete a series of operations such as data collection, standardization processing, scoring calculation, trend prediction, weight optimization, and traffic adjustment, reducing the need for manual intervention and improving the overall effectiveness of SMS marketing.

[0413] In summary, the SMS marketing performance optimization method and system based on multi-channel dynamic weighting provided by this invention achieves intelligent allocation and performance optimization of SMS marketing channel resources through real-time dynamic weight allocation mechanisms, multi-dimensional channel evaluation models, machine learning time-series prediction capabilities, A / B testing automation frameworks, and anomaly detection and emergency handling mechanisms. This method overcomes the limitations of traditional static configuration, improves the accuracy and conversion efficiency of SMS marketing, and ensures stable system operation even in the face of channel failures.

[0414] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-channel dynamic weight-based SMS marketing effect optimization method, characterized in that, The method comprises the following steps: acquiring real-time performance data of a plurality of short message channels; standardizing the real-time performance data, mapping the reach rate, conversion rate, complaint rate and delay indicators to the [0, 1] interval by using the Z-score standardization method to obtain standardized performance indicators; based on the standardized performance indicators, calculating the comprehensive performance score of each channel by weighted summation, using the scoring formula: generating a channel performance benchmark score, wherein the reach weight , the conversion weight , the complaint rate weight , the delay indicator weight satisfies establishing a weight configuration scheme; collecting multi-dimensional performance data of each channel in the past 24 hours, analyzing the time sequence characteristics of historical conversion data by using an LSTM neural network, using a three-layer LSTM network structure to learn the time sequence pattern, and predicting the conversion rate trend change of each channel in the next 2-4 hours; Based on the conversion rate trend change and the channel performance benchmark score, a gradient optimization method is used to solve an objective function of maximizing the overall ROI while satisfying constraint conditions a weight distribution scheme, wherein the gradient optimization method is used to solve an objective function of maximizing the overall ROI while satisfying constraint conditions wherein n is the number of channels, represents the weight proportion allocated to the i-th short message channel, the predicted conversion rate i represents the average conversion rate of the i-th short message channel in the next 2-4 hours, the weight coefficient i represents the adjustment parameter related to the channel characteristics in the i-th short message channel, and the weight coefficient i = base coefficient i × quality factor i × cost factor i wherein the quality factor i is calculated based on the channel performance benchmark score of the i-th short message channel, the base coefficient i is used to reflect the importance of the i-th short message channel, and the cost factor i is inversely proportional to the sending cost of the i-th short message channel; based on the objective function, solving by using the Adam optimization algorithm, setting the learning rate to 0.01 and the maximum number of iterations to 100, and calculating the weight distribution vector that converges to the optimal solution; acquiring the weight distribution vector, comparing and analyzing the current weight configuration, and outputting the optimal distribution ratio; according to the optimal distribution ratio, real-time adjusting the short message queue distribution strategy, re-distributing the to-be-sent short messages to each channel queue, and simultaneously starting the weight configuration smooth transition mechanism to realize dynamic optimization of short message marketing effect.

2. The method of claim 1, wherein, The standardization processing of the real-time performance data, the Z-score standardization method is used to map the reach rate, conversion rate, complaint rate and delay indicators to the [0, 1] interval to obtain standardized performance indicators, which comprises: collecting the original data of the reach rate, conversion rate, complaint rate and delay indicators of each short message channel, calculating the mean μ and standard deviation σ of each indicator in the historical data, and establishing the data distribution baseline; Based on the data distribution baseline, a Z-score normalization formula is applied where X represents the original data value; normalizing each indicator to eliminate the influence of different indicator dimension differences and obtain the normalization transformation result; interval mapping the normalization transformation result, mapping the Z value to the [0, 1] standard interval by using the Sigmoid function to obtain the standardized performance indicators.

3. The method of claim 1, wherein, The collection of multi-dimensional performance data of each channel in the past 24 hours, the analysis of the time sequence characteristics of historical conversion data by using an LSTM neural network, the use of a three-layer LSTM network structure to learn the time sequence pattern, and the prediction of the conversion rate trend change of each channel in the next 2-4 hours, which comprises: based on the multi-dimensional performance data, calculating the 3-hour, 6-hour and 24-hour moving average of the conversion rate, constructing the first-order difference and second-order difference trend characteristics of the conversion rate, and forming the input feature vector of the prediction model; based on the input feature vector of the prediction model, designing a three-layer LSTM network structure, wherein the input layer receives 24 time window feature vectors, the hidden layer captures long and short term dependencies through 128 LSTM units, and a time sequence prediction model is established; deploying the time sequence prediction model as an online inference service, combining the standardized performance indicators, and predicting the conversion rate trend change of each channel in the next 2-4 hours.

4. The method of claim 1, wherein, The smooth transition mechanism of the weight configuration comprises: acquiring the optimal distribution ratio, using an exponential decay smoothing algorithm, setting the smoothing coefficient α to 0.3, and establishing a weight smoothing transition formula; Based on the weight smoothing transition formula, the smoothed weight = a x updated weight + (1-a) x current weight is calculated to generate a gradual weight adjustment sequence; According to the gradual weight adjustment sequence, the weight switching operation is performed in steps to avoid the impact of sudden weight changes on system stability.

5. The method of claim 1, wherein, It also includes an A / B testing framework for verifying the effectiveness of new channels: Randomly select no more than 5% of user traffic as test samples, and split the traffic by user ID hash modulo to allocate test traffic to the channel to be tested, and establish an experimental group and a control group; Collect conversion data from the experimental group and the control group, calculate the confidence interval of conversion rate based on Beta-Binomial conjugate prior distribution, perform Bayesian statistical test, and obtain Bayesian statistical test results; Based on the Bayesian statistical test results, when the conversion rate difference between the experimental group and the control group reaches statistical significance and the relative improvement exceeds 5%, the new channel that passes the verification is automatically included in the weight distribution pool, and an updated channel weight distribution pool is obtained.

6. The method of claim 1, wherein, It also includes deep mining of channel performance based on distributed asynchronous exploration: Get the current channel configuration state of the optimal distribution ratio as the starting point of exploration, and use channel weight, sending period, and content template as exploration dimensions to build a multi-dimensional decision tree structure containing n channel nodes and depth D, and generate a channel configuration exploration tree; Based on the channel configuration exploration tree, generate k configuration variation branches in parallel, assign each branch to a distributed computing node for asynchronous execution, and use small-scale A / B testing to verify the conversion effect of each configuration scheme, and collect key indicator data including conversion rate and ROI; According to the key indicator data, a regret value of each configuration scheme is calculated, inefficient branches are pruned based on a regret minimization principle, optimal branches are retained to continue deep exploration, and the global optimal configuration is converged within a maximum of n times of moving operations; wherein n is the number of channels, k is the number of parallel branches, and D is the exploration depth.

7. The method of claim 6, wherein, The generation of the channel configuration exploration tree includes: Set the current channel configuration state of the optimal distribution ratio as the root node, which contains the complete parameter set of weight distribution, sending strategy, and content template of all channels, and establish the starting state of exploration; Based on the starting state of exploration, fine-tune each configuration parameter to generate multiple sub-node branches, where each sub-node represents a specific variation of a parameter, forming a configuration variation space; Expand the configuration variation space to depth D according to the depth-first search strategy, and each leaf node corresponds to a specific configuration scheme to generate the channel configuration exploration tree.

8. The method of claim 1, wherein, It also includes channel stability prediction based on a two-dimensional threshold cellular automaton: Get the attribute information of all SMS channels, including operator type and geographic area, map each channel to a two-dimensional grid structure, and each grid cell contains a state vector of performance score, load rate, failure probability, and neighborhood influence degree, and establish a channel network topology model; Based on the channel network topology model, design threshold conversion rules: when the channel performance benchmark score is less than 60 and the continuous low score time exceeds 1 hour, it is determined as a warning state; when the number of failed channels in the neighborhood is greater than or equal to 3 and the current load rate exceeds 80%, the failure probability is increased, and the cellular automaton state is updated; According to the cellular automaton state update, 100 time steps are iteratively evolved to simulate the propagation and diffusion process of channel failure, calculate the overall system stability index, and predict the stability trend of the channel network in the next 2-4 hours.

9. The method of claim 8, wherein, The execution cellular automaton state update comprises: Obtain the state vector of each channel at the current time, and determine the state according to the performance threshold rule, neighborhood propagation rule and cascade failure rule, and calculate the state probability distribution of each channel at the next time; Based on the state probability distribution, the state values of all channels are updated in parallel, the neighborhood influence degree and the failure propagation coefficient are calculated synchronously, the diffusion process of the failure in the channel network is simulated, and the failure diffusion process is obtained; Quantitative evaluation is performed on the failure diffusion process, and the stability indexes including the proportion of healthy channels and the size of the largest connected component are calculated to form the overall system stability score.

10. The method of claim 1, wherein, It also includes channel routing optimization based on fast approximate minimum spanning tree: Extract user features including geographical location, device type, historical behavior preference, active period, and channel features including operator type, coverage area, delay characteristics, and cost structure, construct a high-dimensional feature space, and generate user-channel combined feature vectors; Based on the user-channel combined feature vector, the FAMST three-stage algorithm is used to construct the optimal connection relationship of channel routing, and through the three stages of approximate nearest neighbor graph construction, neural network component connection discovery, and iterative edge refinement, the personalized routing network topology is established; According to the personalized routing network topology, the shortest distance and matching degree to each channel are calculated for each sending task, and the optimal channel is selected by combining the optimal distribution ratio to realize user-level precise channel routing optimization.

11. The method of claim 10, wherein, The establishment of the personalized routing network topology comprises: Based on the user-channel combined feature vector, an approximate nearest neighbor graph is constructed using Locality Sensitive Hashing, 10 most similar user-channel combinations are calculated, and an initial connection network is established; The initial connection network is input into a neural network model, node features are extracted by an encoder, and connection probabilities between node pairs are calculated using a connection predictor to discover potential high-value connection relationships; Iterative edge refinement is performed on the potential high-value connection relationships, the lowest 5% edges are removed, network connectivity is checked and key connections are added, and after at most 10 iterations of optimization, the personalized routing network topology is generated.

Images