Construction method of cigarette brand analysis model

Abstract

The invention provides a method for constructing a cigarette brand analysis model, which comprises the following steps of: acquiring combustion speed data and sales data of different brands of cigarettes, and constructing an associated data set of combustion speed and consumption frequency through data preprocessing; combining the prediction result of the combustion speed and the influence coefficient to establish a cigarette demand prediction model, and judging the demand trend and purchase quantity distribution of cigarettes with different combustion characteristics in a future period of time; based on the prediction result of the demand trend and purchase quantity distribution of the cigarettes in a future period of time, adopting a multi-objective optimization algorithm to construct a purchase decision optimization model, and solving the model to obtain the optimal purchase quantity and purchase opportunity of each type of cigarettes; and regularly monitoring the consumption behavior change and competing product dynamic state of the cigarette market, updating the data set and adjusting the prediction model.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for constructing a cigarette brand analysis model. Background Art

[0002] The burning speed of cigarettes is one of the important factors that affect consumer purchasing behavior. There are differences in the burning speed of different brands of cigarettes, and this difference will lead to changes in the frequency and quantity of consumer purchases. In order to better understand this relationship, it is necessary to establish a prediction model that comprehensively considers multiple variables such as cigarette burning speed, consumption frequency, and purchase quantity, and introduces time series analysis methods to capture the dynamic changes of consumer behavior in different time periods. However, in the modeling process, how to accurately measure the burning speed of cigarettes is a technical problem. The burning speed is affected by multiple factors such as the physical properties of cigarettes and environmental conditions, and the differences between different brands are small, so traditional measurement methods are difficult to achieve accurate quantification. At the same time, consumer purchasing behavior also has a certain degree of randomness and uncertainty. How to reasonably characterize this uncertainty in the model and establish an association with the burning speed is another thorny problem. In other words, when establishing a prediction model, how to solve the problem of accurate measurement of cigarette burning speed and uncertainty characterization of consumer behavior while ensuring the accuracy and generalization ability of the model is a technical problem that needs to be solved urgently. Summary of the invention

[0003] The present invention provides a method for constructing a cigarette brand analysis model, which mainly includes:

[0004] Obtain the burning speed data and sales data of different brands of cigarettes, and build a data set related to burning speed and consumption frequency through data preprocessing;

[0005] Based on the correlation data set of burning speed and consumption frequency, a time series prediction model of burning speed and consumption frequency was established by using time series analysis method. The optimal prediction model parameters were obtained through model training and verification. On the basis of time series analysis, the key characteristics that affect the burning speed of cigarettes, including the length, diameter and filling of cigarettes, were judged, and the characteristic indicators that were significantly correlated with the burning speed were screened out.

[0006] According to the selected key combustion characteristic indicators, the cigarette samples are divided into different levels according to different burning speeds. The established time series prediction model is used to analyze the influence of cigarettes of different burning speed levels on consumer purchasing behavior. Through multiple regression analysis, the influence coefficient of each combustion characteristic indicator on the purchase quantity is obtained.

[0007] Combined with the prediction results and influence coefficient of burning speed, a cigarette demand prediction model is established to determine the demand trend and purchase quantity distribution of cigarettes with different burning characteristics in the future.

[0008] Based on the forecast results of demand trend and purchase quantity distribution of cigarettes in the future, a multi-objective optimization algorithm is used to build a purchase decision optimization model. By solving the model, the optimal purchase quantity and purchase timing of various types of cigarettes are obtained.

[0009] Compare and analyze the optimized procurement decision plan with the current procurement strategy, evaluate the effect of procurement optimization, and dynamically adjust the parameters of the prediction model and optimization model based on the feedback results to continuously improve the scientificity and effectiveness of the procurement strategy;

[0010] Regularly monitor changes in consumer behavior and competitive product dynamics in the cigarette market, update data sets and adjust forecasting models.

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The invention discloses a method for constructing a cigarette brand analysis model. The method collects the burning speed and sales data of cigarettes of different brands, constructs a related data set, and uses an ARIMA model to perform time series prediction. At the same time, the key characteristic indicators that affect the burning speed are screened out through correlation analysis, and the cigarette samples are divided into different burning speed levels. Combined with multivariate regression analysis, the influence coefficient of each burning characteristic indicator on the purchase quantity is obtained, and a demand prediction model is established. Finally, a multi-objective optimization algorithm is used to construct a procurement decision optimization model to obtain the optimal procurement plan. The present invention organically combines cigarette characteristics, consumer behavior and market demand through the key factor of burning speed, realizes accurate prediction of cigarette market demand and scientific optimization of procurement decisions, and provides an effective decision support tool for cigarette companies. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The present invention is a flowchart of a method for constructing a cigarette brand analysis model.

[0014] Figure 2 A schematic diagram of a method for constructing a cigarette brand analysis model of the present invention. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] like Figure 1-2 In this embodiment, a method for constructing a cigarette brand analysis model may specifically include:

[0017] Step S101, obtaining burning speed data and sales data of different brands of cigarettes, and constructing a data set associated with burning speed and consumption frequency through data preprocessing.

[0018] Obtain the burning speed data and sales data of different brands of cigarettes and preprocess the data; perform cluster analysis on the feature vectors through the DBSCAN algorithm based on the preprocessed data set; perform correlation analysis between the burning speed and the consumption frequency based on the customer group portrait and feature vector; calculate the Pearson correlation coefficient between the burning speed and the consumption frequency; if the correlation coefficient is greater than the preset threshold, it is determined that there is a significant correlation between the burning speed and the consumption frequency; perform cross-validation on the data set to obtain the stability assessment value of the correlation analysis results; if the stability assessment value is higher than the preset standard, it is determined that the correlation analysis results have high reliability; based on the correlation analysis results, determine the degree of influence of the burning speed on the consumption frequency; based on the degree of influence, generate a correlation coefficient matrix between the burning speed and the consumption frequency; through the correlation coefficient matrix, determine the relationship between the burning speed of different brands of cigarettes and consumer purchasing behavior.

[0019] Specifically, the burning speed data and sales data of different brands of cigarettes are obtained, and the pandas library is used for data preprocessing and feature engineering. The data is processed according to the sampling period, the missing values ​​are filled by the fillna method, and the outliers are identified and deleted by the quantile method. The burning speed and sales volume are standardized to obtain a standardized data set. The groupby and agg methods are used to calculate the average burning time and monthly sales volume of each brand of cigarettes, and the cigarettes are classified according to the price range. The sales volume share of different price ranges is counted to generate feature vectors. For the preprocessed data set, the DBSCAN algorithm is used to perform cluster analysis on the feature vectors. Clustering is implemented using the DBSCAN function in the sklearn library, and appropriate eps and min_samples parameters are set. Different target customer groups are divided according to the clustering results. The average consumption frequency of each customer group is calculated, and the burning speed range preferred by the customer group is determined by counting the burning speed distribution within each customer group. Based on this information, a customer group portrait is constructed, including features such as consumption frequency and burning speed preference. Based on the customer group portrait and feature vector, the correlation analysis between burning speed and consumption frequency is performed. The Pearson correlation coefficient between burning speed and consumption frequency is calculated using the corr method of pandas. A heat map is drawn using the seaborn library to intuitively display the correlation between burning speed and consumption frequency. The correlation coefficients are calculated for different product types and the correlation degree of each type of product is compared. To verify the stability of the association analysis results, the data set is cross-validated. The data set is divided into a training set and a test set, and the association analysis process is repeated. The correlation coefficients obtained from different subsets are compared to evaluate the consistency of the results. The importance score of each feature is output, including the influence of brand, price range, and average burning time on consumption frequency. A correlation coefficient matrix is ​​generated to show the relationship between each feature. After obtaining the burning speed and sales data of different brands of cigarettes, the pandas library is used for data preprocessing and feature engineering. First, the raw data with a sampling period of daily is processed, and the missing burning speed values ​​are filled with the average value of the brand. The outliers in the sales data, such as daily sales exceeding 10,000, are deleted. Then, the burning speed and sales volume are standardized to have a mean of 0 and a standard deviation of 1. Next, calculate the average burning time of each brand of cigarettes, such as 5.2 minutes / stick for brand A and monthly sales, such as 500,000 sticks / month for brand B. According to the price, cigarettes are divided into three ranges: low, medium and high, and the sales share of each range is counted, such as 30% in the low-price area, 50% in the medium-price area, and 20% in the high-price area. These features constitute the basic data set for subsequent analysis. Use the DBSCAN algorithm to perform cluster analysis on the feature vectors, setting the eps parameter to 0.5 and the min_samples parameter to 5. The clustering results show that the three main customer groups are fast smokers, slow smokers and medium smokers.The average consumption frequency of each customer group is calculated, such as 20 cigarettes / day for the fast smoking group, 10 cigarettes / day for the slow smoking group, and 15 cigarettes / day for the medium smoking group. By analyzing the burning speed distribution within each customer group, the preferred burning speed range of each group is determined, such as 3-4 minutes / stick for the fast smoking group, 6-7 minutes / stick for the slow smoking group, and 4.5-5.5 minutes / stick for the medium smoking group. This information constitutes a detailed customer group portrait. Based on the customer group portrait and feature vector, the Pearson correlation coefficient between the burning speed and the consumption frequency is calculated. The results show that the burning speed and the consumption frequency are moderately negatively correlated overall, with a correlation coefficient of -0.6. The heat map drawn intuitively shows this relationship, and the darker the color, the stronger the correlation. The correlation coefficients are calculated for different product types, and it is found that the burning speed of high-priced products has a weak correlation with the consumption frequency, with a correlation coefficient of -0.3, while the low-priced products have a strong correlation, with a correlation coefficient of -0.8. This finding provides a basis for the marketing strategies of products at different price points. To verify the stability of the association analysis results, the data set was divided into a training set and a test set in a ratio of 8:2, and the association analysis process was repeated 5 times. Comparing the correlation coefficients obtained from different subsets, it was found that the results fluctuated within the range of ±0.05, indicating that the analysis results have good stability. The output feature importance score shows that the price range has the greatest impact on consumption frequency (score 0.8), followed by brand (score 0.6) and average burning time (score 0.4). The generated correlation coefficient matrix further shows the relationship between the features, providing data support for formulating product strategies.

[0020] Step S102, based on the associated data set of burning speed and consumption frequency, a time series prediction model of burning speed and consumption frequency is established by using a time series analysis method, and the optimal prediction model parameters are obtained through model training and verification. On the basis of the time series analysis, the key characteristics that affect the burning speed of cigarettes are judged, including the length, diameter, and filler of the cigarette, and the characteristic indicators that are significantly correlated with the burning speed are screened out.

[0021] The augmented Dickey-Fuller test was performed on the time series data. If the p-value of the test result was greater than the preset threshold, it was converted into a stationary series through a difference operation until the p-value was less than the preset threshold. The autocorrelation function and partial autocorrelation function were calculated based on the stationary series to determine the initial order of the ARIMA model. The grid search method was used to traverse the ARIMA model parameter combination, and the optimal model parameters were selected according to the Akaike information criterion to obtain the time series prediction model of burning speed and consumption frequency. The time series prediction model was applied to the test data set and the rolling prediction method was used for prediction. The length, diameter, and filler characteristic data of cigarettes were extracted from the cigarette production parameter database, and the correlation coefficients between each characteristic and the burning speed were calculated in combination with the burning speed data. The principal component analysis method was used to reduce the dimensionality of highly correlated characteristics, and the principal components with a cumulative explained variance ratio greater than the preset threshold were selected. According to the loading matrix of the principal component, the characteristic indicators significantly correlated with the burning speed were screened out.

[0022] Specifically, according to the associated data set of burning speed and consumption frequency, the augmented Dickey-Fuller test is performed on the time series data to determine the stationarity of the sequence. If the test result shows that the p value is greater than 0.05, it is converted into a stationary sequence through a difference operation, and the difference order gradually increases from 1 until the p value is less than 0.05. The autocorrelation function and partial autocorrelation function are calculated to determine the initial order of the ARIMA model. Through the grid search method, the ARIMA model parameter combination is traversed, and the range of p and q is set to 0 to 5, and the range of d is 0 to 2. The model parameters are estimated using the maximum likelihood estimation method, and the optimal model parameters are selected according to the Akaike information criterion to obtain the time series prediction model of burning speed and consumption frequency. The rolling prediction method is used to apply the prediction model to the test data set. The rolling window size is set to 80% of the original training set, and the prediction step is 1. After each prediction, the actual value is added to the training set and the model is refitted. The mean absolute percentage error is calculated. If the error exceeds the preset threshold of 5%, the model parameters are returned to readjust until a prediction model that meets the accuracy requirements is obtained. The time series prediction results are combined with the cigarette characteristic data. The length, diameter, filling and other characteristic data of cigarettes were extracted from the cigarette production parameter database. Combined with the burning speed data, the correlation between each characteristic and the burning speed was calculated by the Pearson correlation coefficient and the Spearman rank correlation coefficient. The principal component analysis method was used to reduce the dimension of highly correlated characteristics, and the principal components with a cumulative explained variance ratio greater than 80% were selected. According to the loading matrix of the principal component, the characteristic indicators significantly correlated with the burning speed were screened out. The augmented Dickey-Fuller test was performed on the correlation data set between the burning speed and the consumption frequency. The results showed that the p value was 0.08, which was greater than the significance level of 0.05, indicating that the sequence was non-stationary. After the first-order difference, the p value dropped to 0.03, meeting the stationary requirement. The autocorrelation function and partial autocorrelation function were calculated, and it was found that the autocorrelation function was truncated at lag 2, and the partial autocorrelation function was truncated at lag 1. The order of the ARIMA model was preliminarily determined to be (1,1,2). Through the grid search method, the ranges of p and q were set to 0 to 5, and the range of d was set to 0 to 2, and a total of 216 parameter combinations were traversed. The model of each group of parameters was estimated using the maximum likelihood estimation method, and the Akaike information criterion value was calculated. Finally, the ARIMA (2,1,2) model with the smallest AIC value was selected as the optimal model, with the AR coefficient being [0.7,-0.3] and the MA coefficient being [0.5,0.2]. The rolling forecast method was adopted, and the rolling window size was set to 240 days. It was assumed that the original training set was 80% of the 300 days, and the forecast step was 1 day. Forecasts were made on the 90-day test set, and the actual values ​​were added to the training set and the model was refitted after each prediction. The calculated mean absolute percentage error was 4.8%, which was less than the preset 5% threshold, and there was no need to return to adjust the parameters. The time series forecast results were combined with the cigarette characteristic data, and characteristic data such as length, diameter, and filler density of 100 types of cigarettes were extracted from the database.The Pearson correlation coefficient was calculated, and it was found that the correlation coefficient between cigarette length and burning speed was 0.75, diameter was -0.62, and filler density was 0.83. Considering the possible nonlinear relationship, the Spearman rank correlation coefficient was calculated, and the results were 0.78, -0.65, and 0.86, respectively, which were close to the Pearson coefficient. Through principal component analysis and dimensionality reduction, the first two principal components were selected, and the cumulative explained variance ratio reached 85%. According to the loading matrix, filler density and cigarette length were determined to be the characteristic indicators most significantly correlated with burning speed.

[0023] Step S103, based on the selected key combustion characteristic indicators, the cigarette samples are divided into different levels according to different burning speeds, and the established time series prediction model is used to analyze the influence of cigarettes of different burning speed levels on consumer purchasing behavior. Through multiple regression analysis, the influence coefficient of each combustion characteristic indicator on the purchase quantity is obtained.

[0024] According to the burning characteristic index of cigarette samples, the hierarchical clustering algorithm is used to divide the cigarette samples into different burning speed levels; according to the burning speed levels obtained by the division, the sales data of cigarettes with different burning speed levels are predicted using the time series prediction model, which specifically includes: extracting the corresponding historical sales data; segmenting the data according to the predefined time window; inputting the segmented data into the time series prediction model to generate the sales forecast value within a certain period of time in the future. The sales trend analysis results are obtained and combined with consumer behavior data, which includes: extracting purchase records from the consumer purchase behavior database; calculating the brand preference index, which is the purchase frequency of a specific brand divided by the total purchase frequency. According to the purchase records and the brand preference index, a consumer behavior feature matrix is ​​constructed, which contains consumer behavior data associated with the cigarette burning speed level. Based on the consumer behavior feature matrix, the generalized additive model is used to establish the relationship between the purchase quantity and the burning characteristic index, which specifically includes: using the spline function to fit the existing nonlinear relationship; removing insignificant variables through the backward screening method; and calculating the influence coefficient of each burning characteristic index on the purchase quantity.

[0025] Specifically, based on the selected key burning characteristic indicators, a hierarchical clustering algorithm is used to divide cigarette samples into different burning speed levels. First, the interquartile range method is used to detect and process outliers. Then, the distance matrix is ​​constructed by calculating the Euclidean distance between samples. Ward's minimum variance method is used for clustering, and the results are visualized through a dendrogram. According to the structure and silhouette coefficient of the dendrogram, the optimal number of clusters is determined to obtain the burning speed range and sample distribution of each level. The established time series prediction model is used to predict the sales data of cigarettes with different burning speed levels. For each burning speed level, the corresponding historical sales data is extracted and the data is segmented according to the predefined time window. The segmented data is input into the time series prediction model to generate sales forecast values ​​for a certain period of time in the future. The prediction accuracy and root mean square error of each level are calculated, and the degree of influence of burning speed on sales trends is quantified by comparing the prediction results of cigarettes of different levels. The sales trend analysis results are combined with consumer behavior data. Purchase records are extracted from the consumer purchase behavior database, including purchase quantity, purchase frequency and price information. The brand preference index was calculated by dividing the purchase frequency of a specific brand by the total purchase frequency. These data were associated with the burning speed grade of cigarettes to construct a matrix of consumer behavior characteristics. The relationship between the sales trend of cigarettes with different burning speed grades and consumer behavior characteristics was explored through cross-analysis. Based on the consumer behavior characteristic matrix, the generalized additive model was used to establish a relationship model between the purchase quantity and the burning characteristic index. The spline function was used to fit the possible nonlinear relationship. The backward screening method was used to remove insignificant variables, and the features with a p-value less than 0.05 were retained. The influence coefficient of each burning characteristic index on the purchase quantity was calculated to obtain the final generalized additive model. The goodness of fit of the model was evaluated using the determination coefficient and the adjusted determination coefficient, and the assumptions of the model were verified by residual analysis. The burning speed grade of cigarette samples was divided into burning speed grades using the hierarchical clustering algorithm. First, the interquartile range method was used to detect outliers, and the upper and lower bounds were set as Q3+1.5IQR and Q1-1.5IQR, respectively. Three abnormal samples were identified and eliminated. Then the Euclidean distances between the remaining 197 samples were calculated to construct a 197x197 distance matrix. Ward's minimum variance method was used for clustering and a dendrogram was generated. By calculating the silhouette coefficients of different cutting points, the optimal number of clusters was determined to be 4, with a silhouette coefficient value of 0.68. The four burning speed levels were obtained: slow 3-4 minutes / stick, medium slow 4-5 minutes / stick, medium fast 5-6 minutes / stick, and fast 6-7 minutes / stick. Time series forecasting was performed on the cigarette sales data of each burning speed level. Taking the medium slow level as an example, the daily sales data of the past 365 days were extracted, and a rolling forecast was performed with a time window of 7 days. The sales volume for the next 30 days was predicted, and the prediction accuracy was calculated to be 92.5%, and the root mean square error was 326 sticks. Comparing the prediction results of the four levels, it was found that the sales trend of medium slow cigarettes was the most stable, with the smallest seasonal fluctuations.The sales trend analysis results were combined with consumer behavior data, and 100,000 purchase records were extracted from the database. The brand preference index was calculated. For example, the index of brand A was 0.35, indicating that 35% of all purchases were for this brand. A consumer behavior feature matrix containing purchase quantity, purchase frequency, price and brand preference index was constructed. Cross-analysis showed that medium and slow cigarettes had the highest brand loyalty, with an average brand preference index of 0.42. Based on the feature matrix, the generalized additive model was used to establish the relationship between purchase quantity and burning characteristic indicators. The cubic spline function was used to fit the nonlinear relationship for each feature, and the variables were screened by the backward screening method. Finally, three significant features of burning speed, cigarette length and filling density were retained (all p values ​​were less than 0.01). The obtained generalized additive model equation showed that the effect of burning speed on purchase quantity showed an inverted U-shaped relationship, and the purchase quantity was the largest in the range of 4.5-5.5 minutes / stick. The adjusted determination coefficient of the model was 0.76, and the residual analysis verified the rationality of the model assumptions.

[0026] Step S104, combining the prediction result of the burning speed and the influence coefficient, establishes a cigarette demand prediction model to determine the demand trend and purchase quantity distribution of cigarettes with different burning characteristics in the future.

[0027] The time series decomposition method is used to process historical sales data to obtain three components: trend, seasonality and random fluctuation. The Monte Carlo simulation method is used to perform random sampling on the time series decomposition results to obtain a cigarette demand forecasting model. After the cigarette demand forecasting model generates multiple sets of forecast results, the k-means clustering algorithm is used to perform market segmentation on the forecast results. Based on the market segmentation results, the demand changes at different price levels are calculated to generate price and demand curves, which are used to determine the demand trend of cigarettes with different burning characteristics. The purchase quantity distribution is estimated through the area under the price and demand curves.

[0028] Specifically, the time series decomposition method is used to decompose historical sales data into three components: trend, seasonality, and random fluctuations. The ARIMA model is combined to process long-term trend changes, and the trend term is differentiated to determine the order of the autoregressive term and the moving average term. The time factor is introduced as an additional independent variable into the generalized additive model to construct a cigarette demand forecasting model. The optimal number of time lag terms is selected by cross-validation to balance the model's fit and generalization ability. The dynamic demand forecasting model is randomly sampled and iterated multiple times by the Monte Carlo simulation method. The number of iterations is set to 10,000, and the model parameters and error terms are randomly sampled. In each iteration, a set of forecast results is generated based on the sampled parameters. All iteration results are summarized, and the mean, standard deviation, and 95% confidence interval of the forecast value are calculated. The simulation results are converted into probability distributions for subsequent market segmentation analysis. The k-means clustering algorithm is used to segment the forecast results. The probability distribution generated by the Monte Carlo simulation is used as the clustering feature, and the number of clusters is set to 3-7. The optimal number of clusters is selected by the silhouette coefficient. Combined with the price elasticity coefficient, the demand forecast for each cluster is adjusted. The demand changes at different price levels are calculated to generate price and demand curves. The demand trend of cigarettes with different burning characteristics in the future period is determined by the slope of the curve, and the purchase quantity distribution is estimated by the area under the curve. Finally, the historical data is used for backtesting, the mean absolute percentage error is calculated, and the prediction accuracy of the model is evaluated. When constructing a cigarette demand forecasting model, the generalized additive model is first used to deal with nonlinear relationships. Taking the burning speed as an example, a cubic spline function with 3 nodes is set, and it is found that the demand is highest in the range of 4.5-5.5 minutes / stick. Through testing, it is found that there is multicollinearity in cigarette length and filler density (variance inflation factor VIF>10), and ridge regression is used for processing, and the regularization parameter λ=0.1. Time series decomposition shows that there is obvious seasonal fluctuation in sales data, with a period of 52 weeks. The ARIMA(2,1,1) model is used to deal with long-term trends, and the difference order d=1 eliminates non-stationarity. The time factor is introduced into the model as a dummy variable, and the optimal lag period is determined to be 4 weeks through 5-fold cross validation. The Monte Carlo simulation was set up for 10,000 iterations, and the model parameters were sampled from a normal distribution, with the standard deviation set to 10% of the estimated value. The simulation results were summarized to obtain the predicted mean and 95% confidence interval, such as the weekly demand forecast for medium-burning speed cigarettes was 100,000 ± 5,000. The predicted distribution was clustered using the k-means algorithm, and the optimal number of clusters was determined to be 4 by the silhouette coefficient. The price elasticity coefficient of each cluster was calculated, such as the elasticity coefficient of fast-burning cigarettes was -1.2, indicating that for every 1% increase in price, the demand decreased by 1.2%. The price and demand curve was generated, and the slope of the curve showed that the upward trend in demand for medium-burning speed cigarettes was the most obvious, and the weekly demand growth rate was expected to be 3% in the next 8 weeks. By estimating the area under the curve, the purchase quantity distribution of slow-burning cigarettes was concentrated in the range of 50,000-70,000 sticks per week.Finally, backtesting using historical data from the past 12 months showed that the model had an average absolute percentage error of 7.5%, validating the reliability of the forecast.

[0029] Step S105, based on the forecast results of demand trend and purchase quantity distribution of cigarettes in the future, a multi-objective optimization algorithm is used to build a purchase decision optimization model, and by solving the model, the optimal purchase quantity and purchase timing of each type of cigarettes are obtained.

[0030] Based on the forecast results of cigarette demand trend and purchase quantity distribution in the future, the Granger causality test method is used to analyze the correlation between different cigarette demands and obtain the cigarette demand correlation matrix; according to the demand forecast results and the correlation matrix, a multi-objective optimization model is established, and the optimization goals are set as minimizing total cost, maximizing service level and minimizing inventory backlog, where service level is defined as order fulfillment rate and inventory backlog is defined as average inventory turnover days; a genetic algorithm is used to solve the multi-objective optimization model, and the genetic algorithm sets the chromosome coding scheme, the purchase quantity is coded with real numbers, the purchase timing is coded with 0-1 integers, and the non-dominated solution set is screened through the Pareto dominance relationship; the hierarchical analysis method is used to determine the weight of each target, and the judgment matrix is ​​constructed. The weight vector is calculated by the eigenvalue method. If the consistency ratio is less than the threshold, the judgment matrix passes the consistency test, and the comprehensive score is calculated, and the solution with the highest score is selected as the final purchase decision plan.

[0031] Specifically, according to the demand forecast results of cigarettes with different burning characteristics, the demand forecast of various cigarettes is processed through the vector autoregression model. The lag order is set to 4, and the optimal model is selected by the AIC criterion. The Granger causality test is used to analyze the correlation between the demands of different cigarettes, and a correlation matrix is ​​constructed. The demand trend for the next 12 weeks is predicted based on the VAR model, and the 95% confidence interval is calculated. The probability distribution of the purchase quantity is obtained by the kernel density estimation method, and the demand forecast curve and purchase quantity distribution diagram of various types of cigarettes are generated. The demand forecast results are converted into the input of the multi-objective optimization model. A multi-objective optimization model is established, and the optimization goals are set to minimize the total cost, maximize the service level, and minimize the inventory backlog. The service level is defined as the order fulfillment rate, and the inventory backlog is defined as the average inventory turnover days. Set inventory cost, out-of-stock cost, procurement cost, and fixed ordering cost as constraints. Construct the objective function:

[0032] C total =C inventory +C shortage +C purchase +C order

[0033] , Ctotal represents the total cost, Cinventory represents the inventory cost, Cshortage represents the out-of-stock cost, Cpurchase represents the purchase cost, and Corder represents the fixed ordering cost. The constraint equations include the inventory balance equation, the capacity constraint, and the non-negative constraint. The genetic algorithm is used to solve the multi-objective optimization model. The chromosome encoding scheme is set, the purchase quantity is encoded with real numbers, and the purchase timing is encoded with 0-1 integers. The population size is set to 100, the crossover rate is 0.8, and the mutation rate is 0.01. The selection operation uses the tournament selection method, the crossover operation uses simulated binary crossover, and the mutation operation uses polynomial mutation. The number of iterations is set to 1000, and the non-dominated solution set is screened through the Pareto dominance relationship. The non-dominated solution set is screened using the Pareto optimal principle. The hierarchical analysis method is used to determine the weights of each target and construct the judgment matrix. The weight vector is calculated by the eigenvalue method and the consistency test is performed. If the consistency ratio CR<0.1, the judgment matrix passes the consistency test. Calculate the comprehensive score and select the solution with the highest score as the final purchase decision plan. The optimal purchase quantity and purchase timing of each type of cigarette are obtained. Sensitivity analysis was performed to evaluate the impact of demand fluctuations and cost changes on decision-making. When constructing a cigarette purchasing decision optimization model, the vector autoregression model was first used to process the demand forecast of multiple cigarettes. Taking three types of cigarettes with burning characteristics (fast, medium, and slow) as an example, the lag order was set to 4, and the optimal model was selected by the AIC criterion, with an AIC value of -1254.6. The Granger causality test showed that there was a significant correlation between the demand for fast-burning cigarettes and the demand for medium-burning cigarettes (p value < 0.01). The VAR (vector autoregression model) model predicts the demand trend in the next 12 weeks. For example, the demand forecast value of fast-burning cigarettes in the first week is 100,000, and the 95% confidence interval is [95,000, 105,000]. The kernel density estimation method obtains the probability distribution of the purchase quantity, and the peak occurs at 98,000. In the multi-objective optimization model, the total cost target f1 weight is set to 0.5, the service level target f2 weight is set to 0.3, and the inventory backlog target f3 weight is set to 0.2. The service level is defined as the order fulfillment rate, with a target value of 98%; the inventory backlog is the average inventory turnover days, with a target value of 15 days. The genetic algorithm parameter settings include population size 100, crossover rate 0.8, mutation rate 0.01, and iteration 1000 times. The final non-dominated solution set contains 27 solutions. The hierarchical analysis method is used to construct the judgment matrix, and the consistency ratio CR=0.04<0.1 passes the consistency test. The solution with the highest comprehensive score is 110,000 fast-burning cigarettes purchased in the second and sixth weeks; 90,000 medium-burning cigarettes purchased in the first and fifth weeks; and 70,000 slow-burning cigarettes purchased in the third week. Sensitivity analysis shows that when demand fluctuates by ±10%, the purchasing decision remains stable.

[0034] Step S106, compare and analyze the optimized procurement decision plan with the current procurement strategy, evaluate the effect of procurement optimization, and dynamically adjust the parameters of the prediction model and optimization model according to the feedback results to continuously improve the scientificity and effectiveness of the procurement strategy.

[0035] The historical data of the optimized purchasing decision plan and the current purchasing strategy were obtained, and a multidimensional data set containing the purchase quantity, purchase timing, inventory level, out-of-stock rate and total cost was constructed through data cleaning and standardization. Based on the multidimensional data set, the optimized purchasing decision plan and the current purchasing strategy were compared and analyzed in terms of various indicators. If the comparative analysis results show that the optimization effect is statistically significant, the degree of improvement of each indicator is calculated. The Bayesian optimization algorithm was used to dynamically adjust the parameters of the cigarette demand forecasting model and the purchasing decision optimization model. The parameter search space was set, and the objective function was defined as the negative value of the prediction error. The next set of parameters was selected by the expected improvement criterion, and the posterior distribution was updated in each iteration to obtain the optimal parameter combination. A purchasing strategy evaluation index system was constructed, and a fuzzy comprehensive evaluation method was used for comprehensive evaluation. A fuzzy evaluation matrix was constructed, and triangular fuzzy numbers were used to represent expert scores. The weight vector was determined by the hierarchical analysis method, and a consistency test was performed. If the consistency ratio was less than the preset threshold, the weight vector was accepted. The comprehensive score was calculated and compared with the preset threshold to judge the scientificity and effectiveness of the purchasing strategy. According to the evaluation results, the parameters of the cigarette demand forecasting model and the purchasing decision optimization model were adjusted.

[0036] Specifically, the historical data of the optimized procurement decision plan and the current procurement strategy are obtained. Through data cleaning and standardization, a multidimensional data set containing procurement quantity, procurement timing, inventory level, out-of-stock rate and total cost is constructed. The time series analysis method is used to process data of different time scales, and seasonal decomposition is used to convert the data to the same time scale. The original indicators are retained to ensure the interpretability of the evaluation indicators. The paired sample T test and Wilcoxon signed rank test methods are used to compare and analyze the optimized procurement decision plan and the current procurement strategy in various indicators. The mean difference and significance level of the two groups of data are calculated to determine whether the effect of procurement optimization is statistically significant. For data that does not conform to the normal distribution, the Wilcoxon test is used. According to the results of the comparative analysis, the improvement degree of each indicator is calculated as the basis for parameter adjustment. The Bayesian optimization algorithm is used to dynamically adjust the parameters of the prediction model and the optimization model. The parameter search space is set, such as the lag order range of the prediction model is 1-10, and the weight range of the optimization model is 0-1. The objective function is defined as the negative value of the prediction error. Gaussian process regression is used as a proxy model, and the next set of parameters is selected by the expected improvement criterion. The number of iterations was set to 100, and the posterior distribution was updated in each iteration, and the optimal parameter combination was finally selected. A procurement strategy evaluation index system was constructed, including cost saving rate, service level improvement, and inventory turnover rate improvement. A fuzzy comprehensive evaluation method was used for comprehensive evaluation. A fuzzy evaluation matrix was constructed, and triangular fuzzy numbers were used to represent expert scores. The weight vector was determined by the hierarchical analysis method and consistency test was performed. If the consistency ratio CR<0.1, the weight vector was accepted. The comprehensive score was calculated and compared with the preset threshold to judge the scientificity and effectiveness of the procurement strategy. According to the evaluation results, the parameters of the prediction model and the optimization model were adjusted to form a feedback loop mechanism to continuously improve the procurement strategy. When comparing and analyzing the optimized procurement decision plan with the current procurement strategy, the historical data was first processed. Taking a certain cigarette brand as an example, 52 weeks of data were collected, including weekly procurement quantity, inventory level, out-of-stock rate, and total cost. Weekly data was converted to monthly data through seasonal decomposition to remove the impact of short-term fluctuations. After data standardization, the mean purchase quantity increased from 95,000 pieces / week to 102,000 pieces / week, the inventory level decreased from an average of 15 days to 12 days, the out-of-stock rate decreased from 3.2% to 1.8%, and the average monthly total cost decreased from 5.2 million yuan to 4.95 million yuan. Paired sample T test was performed on these indicators, and the p values ​​were all less than 0.05, indicating that the differences were statistically significant. For the out-of-stock rate data that did not conform to the normal distribution, the Wilcoxon signed rank test was used, and the p value was 0.03, which also supported the significant optimization effect. Based on these results, it was calculated that the purchase quantity increased by 7.4%, the inventory turnover rate increased by 20%, the out-of-stock rate decreased by 43.8%, and the cost savings were 4.8%. These improvement degrees were used as the basis for parameter adjustment and input into the Bayesian optimization algorithm.In the parameter optimization process, the lag order search range of the prediction model is set to 1-10, and the cost weight range of the optimization model is set to 0.4-0.6. Using the negative value of the prediction error as the objective function, after 100 iterations, the optimal lag order is finally determined to be 4 and the cost weight is 0.52. In the evaluation stage, a fuzzy evaluation matrix is ​​constructed, and the experts' scores for cost saving rate, service level improvement and inventory turnover rate improvement are (0.8, 0.9, 1.0), (0.7, 0.8, 0.9) and (0.8, 0.9, 0.95) respectively. The weight vector (0.4, 0.3, 0.3) is obtained by the hierarchical analysis method, and the consistency ratio CR = 0.02 < 0.1, and the weight is accepted. The calculated comprehensive score is 0.86, which is higher than the preset threshold of 0.8, indicating that the optimized procurement strategy is highly scientific and effective. Based on the evaluation results, the parameters of the prediction model are further fine-tuned to form a closed-loop mechanism for continuous improvement.

[0037] Step S107, regularly monitor changes in consumer behavior and competitive product dynamics in the cigarette market, update the data set and adjust the prediction model.

[0038] Receive consumer comments, competitive product information and market dynamics obtained in real time from public channels by data collection tools, which constitute the original data; execute a data cleaning process based on the original data, which includes format unification, deduplication and missing value processing; perform natural language processing on the data processed by the data cleaning process to obtain a structured market monitoring data set, which includes keywords and sentiment tendencies; use online learning methods to update the cigarette demand forecasting model, and integrate the market monitoring data set with the historical data set through online learning methods; construct an early warning indicator system based on the updated results of the forecasting model, which includes thresholds for changes in consumption trends, adjustments to competitive product strategies and abnormal market fluctuations; if an indicator in the early warning indicator system exceeds a preset threshold, a warning signal is triggered; adjust the data collection frequency and scope according to the level of the warning signal.

[0039] Specifically, a cigarette market monitoring system was built to obtain consumer comments, competitive product information and market dynamics from public channels in real time through data collection tools. Data source compliance checks were performed to ensure that data collection complied with legal and privacy regulations. The heterogeneity of data from different sources was handled using a data cleaning process, including format unification, deduplication and missing value processing. Keywords and sentiment tendencies were extracted using a natural language processing algorithm to form a structured market monitoring data set. Based on the market monitoring data set, the LDA (latent Dirichlet allocation) algorithm was used to identify the main features of changes in consumer behavior and competitive product dynamics. Consumers were divided into different groups through K-means clustering analysis, and brand preferences, price sensitivity and product demand changes of each group were calculated. The results of consumer group analysis were converted into feature vectors as input for updating the prediction model. The prediction model was updated using an online learning algorithm to integrate the newly added market monitoring data with the historical data set. Model parameters and feature weights were dynamically adjusted to incorporate the results of consumer group analysis. Incremental training was performed using the sliding window method, with the window size set to the most recent 30 days of data. Model performance was evaluated using k-fold cross validation, and the model retraining mechanism was triggered if the prediction error exceeded the preset threshold of 5%. Build an early warning indicator system and set thresholds for changes in consumption trends, adjustments in competitive product strategies, and abnormal market fluctuations. Use the ARIMA model combined with the moving window method to detect time series anomalies, with the window size set to 90 days. Identify abnormal points and generate early warning signals. Automatically adjust the data collection frequency and scope according to the warning level, increase the collection frequency by 10% for low-level warnings, and expand the collection scope by 20% for high-level warnings. Implement data quality control mechanisms, including data integrity checks and outlier identification, to ensure the accuracy and representativeness of collected data. When building a cigarette market monitoring system, the data collection tool obtains approximately 500,000 consumer reviews and 1,000 competitive product information from 10 major e-commerce platforms and 5 social media sites every day. The data source compliance check eliminates 15% of non-public data. The data cleaning process converts 90% of unstructured data into a unified JSON format, with a deduplication rate of 5%, and missing values ​​are processed by mean filling for 3% of the data. The natural language processing algorithm extracts 500 keywords with an accuracy rate of 85% in sentiment tendency. The LDA algorithm was used to identify eight major consumer behavior themes, such as "taste preference" and "price sensitivity". K-means clustering divided consumers into four groups, and calculated the brand preference of each group, such as Group A's preference for brand X was 0.75, price sensitivity, such as Group B's price elasticity of -1.2, and product demand changes, such as Group C's demand for mint products increased by 8% month-on-month. The online learning algorithm updated the prediction model daily, with the sliding window size set to 30 days and the weight of the latest 7-day data being 0.6. The 5-fold cross validation showed that the model accuracy was 92%, which was higher than the preset threshold of 90%. The ARIMA model used a 90-day moving window to detect three anomalies, generating two low-level warnings and one high-level warning.The frequency of data collection increased by 10% and 20% accordingly. The data quality control mechanism identified 2% of outliers and removed 1.5% of the data after manual review. The accuracy of the final collected data reached 98%, representing the top 80% of brands in market share.

[0040] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the concept of the present application. For example, the above features are replaced with the technical features with similar functions disclosed in the present application (but not limited to) to form a technical solution.

Claims

1. A method for constructing a cigarette brand analysis model, characterized in that: The method comprises: Obtain the burning speed data and sales data of different brands of cigarettes, and build a data set related to burning speed and consumption frequency through data preprocessing; Based on the correlation data set of burning speed and consumption frequency, a time series prediction model of burning speed and consumption frequency was established by using time series analysis method. The optimal prediction model parameters were obtained through model training and verification. On the basis of time series analysis, the key characteristics that affect the burning speed of cigarettes, including the length, diameter and filling of cigarettes, were judged, and the characteristic indicators that were significantly correlated with the burning speed were screened out. According to the selected key combustion characteristic indicators, the cigarette samples are divided into different levels according to different burning speeds. The established time series prediction model is used to analyze the influence of cigarettes of different burning speed levels on consumer purchasing behavior. Through multiple regression analysis, the influence coefficient of each combustion characteristic indicator on the purchase quantity is obtained. Combined with the prediction results and influence coefficient of burning speed, a cigarette demand prediction model is established to determine the demand trend and purchase quantity distribution of cigarettes with different burning characteristics in the future. Based on the forecast results of demand trend and purchase quantity distribution of cigarettes in the future, a multi-objective optimization algorithm is used to build a purchase decision optimization model. By solving the model, the optimal purchase quantity and purchase timing of various types of cigarettes are obtained. Compare and analyze the optimized procurement decision plan with the current procurement strategy, evaluate the effect of procurement optimization, and dynamically adjust the parameters of the prediction model and optimization model based on the feedback results to continuously improve the scientificity and effectiveness of the procurement strategy; Regularly monitor changes in consumer behavior and competitive product dynamics in the cigarette market, update data sets and adjust forecasting models.

2. The method according to claim 1, characterized in that The method of obtaining burning speed data and sales data of different brands of cigarettes and constructing a data set of association between burning speed and consumption frequency through data preprocessing includes: Obtain the burning speed data and sales data of different brands of cigarettes and pre-process the data; Based on the preprocessed data set, cluster analysis is performed on the feature vectors using the DBSCAN algorithm; Conduct correlation analysis between burning speed and consumption frequency based on customer profiles and feature vectors; Pearson correlation coefficients between burning rate and consumption frequency were calculated; If the correlation coefficient is greater than the preset threshold, it is determined that there is a significant correlation between the burning speed and the consumption frequency; Perform cross validation on the data set to obtain a stability assessment value for the association analysis results; If the stability assessment value is higher than the preset standard, it is determined that the association analysis result has high reliability; Based on the results of the correlation analysis, determine the degree of influence of the burning speed on the consumption frequency; According to the degree of influence, a correlation coefficient matrix between burning speed and consumption frequency is generated; The relationship between the burning speed of different brands of cigarettes and consumer purchasing behavior is determined through the correlation coefficient matrix.

3. The method according to claim 1, characterized in that According to the associated data set of burning speed and consumption frequency, a time series prediction model of burning speed and consumption frequency is established by using a time series analysis method. The optimal prediction model parameters are obtained through model training and verification. On the basis of time series analysis, the key characteristics that affect the burning speed of cigarettes, including the length, diameter, and filler of cigarettes, are judged, and characteristic indicators that are significantly related to the burning speed are screened out, including: Perform an augmented Dickey-Fuller test on the time series data. If the p-value of the test result is greater than the preset threshold, it is converted into a stationary series through a difference operation until the p-value is less than the preset threshold. Calculate the autocorrelation function and partial autocorrelation function based on the stationary sequence and determine the initial order of the ARIMA model; The grid search method is used to traverse the ARIMA model parameter combination, and the optimal model parameters are selected according to the Akaike Information Criterion to obtain the time series prediction model of burning speed and consumption frequency. The time series forecasting model is applied to the test data set and the rolling forecast method is used for forecasting; Extract the length, diameter and filler characteristic data of cigarettes from the cigarette production parameter database, combine with the burning speed data, and calculate the correlation coefficient between each characteristic and the burning speed; The principal component analysis method is used to reduce the dimension of highly correlated features, and the principal components with a cumulative explained variance ratio greater than a preset threshold are selected; According to the loading matrix of the principal components, characteristic indicators significantly related to the burning rate are screened out.

4. The method according to claim 1, characterized in that: According to the selected key combustion characteristic indicators, the cigarette samples are divided into different levels according to different burning speeds, and the established time series prediction model is used to analyze the influence of cigarettes of different burning speed levels on consumer purchasing behavior. Through multiple regression analysis, the influence coefficient of each combustion characteristic indicator on the purchase quantity is obtained, including: According to the burning characteristic indexes of cigarette samples, the hierarchical clustering algorithm is used to divide the cigarette samples into different burning speed levels. According to the burning speed levels obtained by the classification, the sales data of cigarettes with different burning speed levels are predicted using a time series prediction model, specifically including: extracting corresponding historical sales data; Split data according to predefined time windows; The segmented data is input into the time series forecasting model to generate sales forecast values ​​for a certain period of time in the future; Obtain sales trend analysis results and combine them with consumer behavior data, including: extracting purchase records from the consumer purchase behavior database; Calculate the brand preference index, which is the purchase frequency of a specific brand divided by the total purchase frequency; Based on the purchase records and brand preference index, a consumer behavior characteristic matrix is ​​constructed, and the consumer behavior characteristic matrix contains consumer behavior data associated with the cigarette burning speed grade; Based on the consumer behavior characteristic matrix, the generalized additive model is used to establish the relationship between the purchase quantity and the combustion characteristic index, specifically including: using spline functions to fit the existing nonlinear relationship; Insignificant variables were removed through backward screening; Calculate the influence coefficient of each combustion characteristic index on the purchase quantity.

5. The method according to claim 1, characterized in that: The prediction results and influence coefficients of the burning speed are combined to establish a cigarette demand prediction model to determine the demand trend and purchase quantity distribution of cigarettes with different burning characteristics in the future, including: The historical sales data is processed by time series decomposition method to obtain three components: trend, seasonality and random fluctuation; Based on the time series decomposition results, random sampling is performed through the Monte Carlo simulation method to obtain the cigarette demand forecasting model; After the cigarette demand forecasting model generates multiple sets of forecast results, the k-means clustering algorithm is used to segment the forecast results into markets; Based on the market segmentation results, the demand changes at different price levels are calculated to generate price and demand curves, which are used to determine the demand trends of cigarettes with different burning characteristics; Estimate the purchase quantity distribution using the area under the price and demand curves.

6. The method according to claim 1, characterized in that The prediction results of demand trend and purchase quantity distribution of cigarettes in the future are based on which a multi-objective optimization algorithm is used to construct a purchase decision optimization model. By solving the model, the optimal purchase quantity and purchase timing of various types of cigarettes are obtained, including: Based on the forecast results of cigarette demand trend and purchase quantity distribution in the future, the Granger causality test method is used to analyze the correlation between different cigarette demands and obtain the cigarette demand correlation matrix. According to the demand forecast results and the correlation matrix, a multi-objective optimization model is established, and the optimization objectives are set as minimizing total cost, maximizing service level and minimizing inventory backlog, where service level is defined as order fulfillment rate and inventory backlog is defined as average inventory turnover days; Genetic algorithm is used to solve the multi-objective optimization model. The genetic algorithm sets the chromosome coding scheme, the purchase quantity is coded with real numbers, the purchase timing is coded with 0-1 integers, and the non-dominated solution set is screened through the Pareto dominance relationship; The analytic hierarchy process is used to determine the weight of each objective, construct a judgment matrix, and calculate the weight vector using the eigenvalue method. If the consistency ratio is less than the threshold, the judgment matrix passes the consistency test, calculates the comprehensive score, and selects the solution with the highest score as the final procurement decision plan.

7. The method according to claim 1, characterized in that The optimized procurement decision plan is compared and analyzed with the current procurement strategy to evaluate the effect of procurement optimization, and the parameters of the prediction model and optimization model are dynamically adjusted according to the feedback results to continuously improve the scientificity and effectiveness of the procurement strategy, including: Obtain the historical data of optimized purchasing decision-making plans and current purchasing strategies, and build a multidimensional data set including purchase quantity, purchase timing, inventory level, out-of-stock rate and total cost through data cleaning and standardization; Based on multidimensional data sets, compare and analyze the optimized procurement decision plan with the current procurement strategy in terms of various indicators; If the comparative analysis results show that the optimization effect is statistically significant, the degree of improvement of each indicator is calculated; The Bayesian optimization algorithm is used to dynamically adjust the parameters of the cigarette demand forecasting model and the purchasing decision optimization model; Set the parameter search space and define the objective function as the negative value of the prediction error; The next set of parameters is selected by the expected improvement criterion, and the posterior distribution is updated in each iteration to obtain the optimal parameter combination; Construct a procurement strategy evaluation index system and use fuzzy comprehensive evaluation method to conduct a comprehensive evaluation; Construct a fuzzy evaluation matrix and use triangular fuzzy numbers to represent expert ratings; The weight vector is determined by the analytic hierarchy process and consistency check is performed; If the consistency ratio is less than the preset threshold, the weight vector is accepted; Calculate the comprehensive score and compare it with the preset threshold to judge the scientificity and effectiveness of the procurement strategy; According to the evaluation results, adjust the parameters of the cigarette demand forecasting model and the purchasing decision optimization model.

8. The method according to claim 1, characterized in that The above-mentioned regular monitoring of changes in consumer behavior and competitive product dynamics in the cigarette market, updating of data sets and adjustment of prediction models include: Receive consumer comments, competitive product information and market trends obtained in real time from public channels by data collection tools. Consumer comments, competitive product information and market trends constitute the original data; Perform data cleaning process based on raw data, including format unification, deduplication and missing value processing; Perform natural language processing on the data processed by the data cleaning process to obtain a structured market monitoring data set, which includes keywords and sentiment tendencies; The cigarette demand forecasting model is updated using online learning methods, which integrate the market monitoring dataset with the historical dataset; Build an early warning indicator system based on the updated results of the forecasting model. The early warning indicator system includes thresholds for changes in consumption trends, adjustments in competitive product strategies, and abnormal market fluctuations. If the indicators in the early warning indicator system exceed the preset threshold, the early warning signal is triggered; Adjust the frequency and scope of data collection according to the level of the early warning signal.

Images