Cigarette consumer analysis method

Abstract

The invention provides a cigarette consumer analysis method, which comprises the following steps of: clustering and grouping consumers according to cigarette attributes in a consumption data set, and dividing the consumers with similar attribute preferences into the same subdivided group, the cigarette attributes comprise cigarette price sensitivity, cigarette taste preference types, cigarette appearance preference styles, cigarette brand loyalty and cigarette purchasing frequency; aiming at each consumer subdivided group obtained by clustering, respectively constructing a cigarette attribute association map of the group, learning low-dimensional vector representation of cigarette attributes in each subdivided group in the map, and measuring correlation among different attributes through distances among vectors; and according to the association rules of the high-nicotine-content preference group and the low-tar-content preference group, performing optimization combination of the association rules, and obtaining a subdivided group portrait model by balancing the relationship among different attributes and considering the optimal association combination of consumer preferences in the subdivided group.

Description

Technical Field

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

[0002] In the analysis of cigarette consumption data, we are faced with the challenge of how to accurately divide consumer groups and mine their attribute preference association patterns. First, the processing of massive cigarette consumption data requires the use of a distributed big data framework for preprocessing and cleaning to construct a structured data set containing attribute preference information such as tobacco leaf moisture content, smoke nicotine content, and smoke tar content. Secondly, based on these attributes and factors such as price sensitivity and taste preference, it is necessary to design a suitable clustering algorithm to divide consumer segments. However, the correlation between different attributes is complex, and traditional clustering methods may find it difficult to capture these subtle relationships. Therefore, it is considered to construct a cigarette attribute association graph and use a graph embedding algorithm to learn a low-dimensional representation between attributes. Although this method can better express attribute relationships, how to efficiently mine attribute preference association rules for different segments on this basis, especially for high nicotine content and low tar content preference groups, is still a problem to be solved. In addition, how to optimize the combination of association rules mined to balance the relationship between different attributes and take into account the preference differences within the group, so as to build an accurate consumer portrait model, is also a technical difficulty that needs in-depth research. Summary of the invention

[0003] The present invention provides a method for analyzing cigarette consumers, which mainly comprises:

[0004] Obtain massive amounts of cigarette consumption data, pre-process the data, and obtain a structured consumption data set through data cleaning and data integration operations. The consumption data set includes cigarette price sensitivity, cigarette flavor preference type, cigarette appearance preference style, cigarette brand loyalty, and cigarette purchase frequency. The cigarette flavor preference type includes the preferred value of tobacco leaf moisture content, the preferred range of smoke nicotine content, and the preferred interval attributes of smoke tar content;

[0005] According to the cigarette attributes in the consumption data set, consumers are clustered and grouped, and consumers with similar attribute preferences are divided into the same segment group. The cigarette attributes include cigarette price sensitivity, cigarette flavor preference type, cigarette appearance preference style, cigarette brand loyalty, and cigarette purchase frequency;

[0006] For each consumer segment group obtained by clustering, a cigarette attribute association map of the group is constructed respectively, and the low-dimensional vector representation of the cigarette attributes within each segment group in the map is learned. The correlation between different attributes is measured by the distance between vectors.

[0007] Based on the association maps of cigarette attributes of each segmented group, the association patterns of cigarette attribute preferences of different segmented groups were found, and the association rules of tobacco leaf moisture content preference, cigarette price sensitivity and appearance preference style of the high nicotine content preference group, and the association rules of smoke nicotine content preference range, flavor preference type and brand loyalty of the low tar content preference group were obtained, forming a representative consumer portrait;

[0008] According to the association rules of the high nicotine preference group and the low tar preference group, the association rules are optimized and combined. By balancing the relationship between different attributes and taking into account the optimal association combination of consumer preferences within the segmented group, a segmented group portrait model is obtained.

[0009] The segmented group portrait model is applied to the actual cigarette product research and development and marketing decision-making. The portrait model is used to predict the preferences of the high-nicotine preference group and the low-tar preference group for new product attribute combinations, and combined with the demographic attribute distribution of different preference groups, product formula adjustment and marketing promotion strategies are formulated.

[0010] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0011] The present invention discloses a method for analyzing cigarette consumers. The method first obtains massive cigarette consumption data, and obtains a structured consumption data set through data preprocessing, including attributes such as price sensitivity and taste preference. Then, the consumers are clustered and grouped, and a cigarette attribute association map of each group is constructed, and a low-dimensional vector representation of the attribute is learned. For the groups with high nicotine content and low tar content preference, the association pattern of their cigarette attribute preference is found to form a representative consumer portrait. By normalizing the attribute data, resampling the samples, etc., the association rule combination is optimized to obtain a segmented group portrait model. Finally, the model is applied to cigarette product research and development and marketing decisions to predict the preferences of different groups for new products, and formulate corresponding product formulas and marketing strategies. The present invention realizes accurate portraits of cigarette consumers and provides effective support for cigarette product innovation and personalized marketing. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The present invention is a flowchart of a method for analyzing cigarette consumers.

[0013] Figure 2 A schematic diagram of a method for analyzing cigarette consumers according to the present invention. DETAILED DESCRIPTION

[0014] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0015] like Figure 1-2 In this embodiment, a cigarette consumer analysis method may specifically include:

[0016] Step S101, obtaining massive cigarette consumption data, preprocessing the data, and obtaining a structured consumption data set through data cleaning and data integration operations. The consumption data set includes cigarette price sensitivity, cigarette flavor preference type, cigarette appearance preference style, cigarette brand loyalty, and cigarette purchase frequency, wherein the cigarette flavor preference type includes the tobacco leaf moisture content preference value, the smoke nicotine content preference range, and the smoke tar content preference interval attributes.

[0017] Obtain consumers' price sensitivity, brand loyalty and purchase frequency data, use K-means clustering algorithm to group consumers, and obtain taste preference feature vectors of different consumer groups; based on the taste preference feature vectors, extract visual features such as color and texture from the cigarette appearance preference style data, and use support vector machine to classify the appearance style; if the support vector machine classification is completed, use content-based recommendation algorithm to generate personalized appearance style and taste combination recommendation list for each consumer group; obtain geographic location distribution data and seasonal consumption trends, and use ARIMA time series analysis method to predict cigarette sales in various regions; based on sales forecast results and promotion activity responsiveness data, construct a random forest model to quantify the nonlinear impact of promotion activities on sales, and determine the optimal promotion strategy combination through feature importance; apply the optimal promotion strategy combination to the personalized recommendation list to obtain the final cigarette product sales plan.

[0018] Specifically, a consumer portrait model was constructed based on cigarette price sensitivity, brand loyalty, and purchase frequency data. The K-means clustering algorithm was used to group consumers. The mean and variance of the preferred value of tobacco moisture content, smoke nicotine content, and smoke tar content for each consumer group were calculated to obtain the taste preference feature vectors of different consumer groups. The price sensitivity index of each consumer group was calculated using the price elasticity model, and the index was added as a feature to the taste preference feature vector. Visual features such as color and texture were extracted from the cigarette appearance preference style data, and the appearance style was classified using a support vector machine. Combined with the consumer's purchase history, a content-based recommendation algorithm was used to generate a personalized appearance style and taste combination recommendation list for each consumer group. Based on the geographic location distribution data and seasonal consumption trends, the ARIMA time series analysis method was used to predict cigarette sales in each region. Combined with the promotion activity responsiveness data, a random forest model was constructed to quantify the nonlinear impact of promotion activities on sales, and the optimal promotion strategy combination was determined by feature importance. According to the characteristics of consumer groups and the preferred types of cigarette flavors, specific attributes such as tobacco leaf moisture content, smoke nicotine content, and smoke tar content are used as feature vectors and input into the content-based recommendation system. The system inputs the demographic characteristics and historical purchase data of consumers and outputs the cigarette flavor type and appearance style that the consumer is most likely to prefer. Based on the prediction results of the recommendation system, the expected demand for each type of cigarette is calculated and compared with the current inventory to generate inventory adjustment suggestions.

[0019] When constructing the consumer portrait model, the K-means clustering algorithm was used to group 100,000 consumers, setting the number of clusters to 5 and the number of iterations to 100, to obtain 5 consumer groups with different characteristics. For each group, the mean and variance of the preferred value of tobacco leaf moisture content were calculated, such as the mean of group A was 13% and the variance was 0.5%; the preferred range of smoke nicotine content was calculated, such as the range of group B was 0.8mg-1.2mg; the preferred interval of smoke tar content was calculated, such as the interval of group C was 8mg-10mg. The price elasticity model was used to calculate the price sensitivity index of each group. For example, the index of group D was -1.5, indicating that it was more sensitive to price. The color and texture features were extracted from the cigarette appearance preference style data, and the support vector machine was used for classification. The kernel function was set to RBF and the penalty parameter C was set to 1.0, and style categories such as "traditional classic" and "fashionable light luxury" were obtained. The content-based recommendation algorithm combines consumer purchase history to generate a recommendation list for each group, such as recommending "fashionable light luxury" style cigarettes with a moisture content of 12%-14%, nicotine content of 0.9mg-1.1mg, and tar content of 9mg-11mg for group E. The ARIMA model is used to predict cigarette sales in each region, setting parameters p=1, d=1, q=1, and predicting sales trends in the next three months. A random forest model is constructed to quantify the impact of promotional activities, setting the number of trees to 100 and the maximum depth to 10. Feature importance analysis shows that price discounts have the greatest impact on sales, with a weight of 0.4. The attributes such as tobacco leaf moisture content, smoke nicotine content, and smoke tar content are used as feature vectors and input into the content-based recommendation system. The system uses cosine similarity to calculate the matching degree between consumer preferences and product features. The threshold is set to 0.8, and the cigarette types with a matching degree greater than the threshold are output. Based on the prediction results of the recommendation system, the expected demand for each type of cigarette is calculated. For example, if the expected demand for a certain brand is 100,000 cartons and the current inventory is 80,000 cartons, an adjustment suggestion is generated to increase the inventory by 20,000 cartons.

[0020] Step S102, clustering consumers according to cigarette attributes in the consumption data set, and dividing consumers with similar attribute preferences into the same segmented group, wherein the cigarette attributes include cigarette price sensitivity, cigarette flavor preference type, cigarette appearance preference style, cigarette brand loyalty, and cigarette purchase frequency.

[0021] According to each attribute data in the consumption data set, standardization processing is performed to obtain standardized attribute data. The principal component analysis method is used to reduce the dimension of the standardized attribute data to obtain the feature vector after dimension reduction. The silhouette coefficient method is used to analyze the feature vector after dimension reduction to determine the optimal number of clusters. According to the optimal number of clusters, K-means cluster analysis is performed on the feature vector after dimension reduction to obtain the segmented group. For the segmented group, the mean and standard deviation of each attribute in the group are calculated to construct a group feature description matrix. Based on the group feature description matrix, a consumer classification model is constructed using a decision tree algorithm to determine the segmented group to which the newly added consumer belongs. If there are missing values ​​in the newly added consumer data, the multiple interpolation method is used to process the missing values. The newly added consumer data is input into the consumer classification model, and the segmented group to which the newly added consumer belongs is determined by comparing the attribute value with the threshold of the decision node.

[0022] Specifically, according to the attributes of cigarette price sensitivity, flavor preference type, appearance preference style, brand loyalty, purchase frequency, tobacco leaf moisture content preference, smoke nicotine content preference, smoke tar content preference, etc. in the consumer data set, all attributes are standardized to ensure that each attribute has the same weight in subsequent analysis. The principal component analysis method is used to reduce the dimension of the standardized attributes. By calculating the eigenvalues ​​and eigenvectors, the principal components with a cumulative variance contribution rate greater than 85% are selected as the eigenvectors after dimensionality reduction. The local outlier factor algorithm is used to identify and remove outliers to improve data quality. The silhouette coefficient method is used to determine the optimal number of clusters, and this number is used as a parameter of the K-means clustering algorithm. Cluster analysis is performed on the eigenvectors after dimensionality reduction, and consumers with similar attribute preferences are divided into the same segment group by calculating the Euclidean distance from the sample point to the cluster center. For each segment group, the mean and standard deviation of each attribute in the group are calculated to construct a group feature description matrix. The significant features of each group were extracted from the matrix. The significant features were defined as features whose attribute values ​​exceeded one standard deviation from the group mean, such as high price sensitivity, strong mint flavor preference, and high loyalty to a specific brand. Based on the feature description matrix of the segmented groups, the decision tree algorithm was used to build a consumer classification model. The missing values ​​in the newly added consumer data were processed using the multiple interpolation method to ensure data integrity. The processed new consumer data were input into the decision tree, and the new consumers were divided into the most matching segmented groups by comparing the attribute values ​​with the thresholds of the decision nodes. The classification accuracy was calculated using the cross-validation method to evaluate the model performance. The cigarette attributes in the consumer data set were standardized, and each attribute value was converted to a standard normal distribution with a mean of 0 and a standard deviation of 1. The principal component analysis method was used for dimensionality reduction, and the eigenvalues ​​and eigenvectors were calculated. The first five principal components were selected as the eigenvectors after dimensionality reduction, and the cumulative variance contribution rate reached 87.3%. The local outlier factor algorithm was used, and the threshold was set to 1.5 to identify and eliminate 2.1% of outliers. By calculating the silhouette coefficient under different numbers of clusters, the optimal number of clusters was determined to be 6, and the silhouette coefficient was 0.68. The cluster number 6 is input into the K-means algorithm, and the feature vector after dimension reduction is clustered to obtain 6 consumer segmentation groups. The mean and standard deviation of each attribute in each group are calculated to construct a 6×8 group feature description matrix. Significant features are defined as features whose attribute values ​​exceed the group mean by one standard deviation. For example, the price sensitivity of group 1 is 1.8, which is 1.2 standard deviations higher than the mean, and the mint flavor preference of group 2 is 2.3, which is 1.5 standard deviations higher than the mean. Based on the feature description matrix, the decision tree algorithm is used to construct a consumer classification model. The depth of the tree is set to 4, and the minimum number of leaf node samples is 50. Multiple interpolation is used to handle missing values ​​in the newly added consumer data, and the number of interpolations is set to 5 to generate a complete data set.The new consumer data is input into the decision tree, and by comparing the attribute value with the threshold of the decision node, if the price sensitivity is >1.5, it is the left subtree, otherwise it is the right subtree, and finally the new consumer is divided into one of the six sub-groups. The classification accuracy is calculated using the 5-fold cross-validation method, and the average accuracy is 85.7% with a standard deviation of 2.1.

[0023] Step S103, for each consumer segment group obtained by clustering, respectively construct a cigarette attribute association map of the group, learn the low-dimensional vector representation of the cigarette attributes within each segment group in the map, and measure the correlation between different attributes by the distance between vectors.

[0024] The maximum spanning tree algorithm is used to construct a cigarette attribute association map, and the connection relationship between attribute nodes is determined by calculating the mutual information value between the attributes. The node2vec algorithm is applied to perform graph embedding based on the cigarette attribute association map to obtain a low-dimensional vector representation of the attribute node. The distance between attribute vectors is calculated using cosine similarity, and the distance value is normalized to between 0 and 1 to obtain the correlation score matrix between different attributes. If the correlation score matrix has been obtained, the hierarchical clustering algorithm is used to group the attributes to obtain attribute association clusters. For the attribute association clusters, the t-SNE algorithm is used to project the high-dimensional attribute vectors onto a two-dimensional plane to generate an attribute relationship visualization diagram, which is used to display the correlation strength and clustering results between different cigarette attributes.

[0025] Specifically, according to the data of each consumer segment group obtained by clustering, the maximum spanning tree algorithm was used to construct the cigarette attribute association map of each group, and the connection relationship between attribute nodes was determined by calculating the mutual information value between attributes. The optimal mutual information threshold was determined by five-fold cross validation, and edge connections were established between attribute nodes with mutual information values ​​higher than the threshold. The node2vec algorithm was applied to the constructed cigarette attribute association map for graph embedding, and each attribute node in the map was mapped to a low-dimensional vector space. The optimal hyperparameters, including vector dimension, window size, and walk length, were determined by grid search to obtain a low-dimensional vector representation of each attribute. The distance between attribute vectors was calculated using cosine similarity, and the calculated distance values ​​were normalized to between 0 and 1 to obtain the correlation score matrix between different attributes. The mean and standard deviation of the correlation score were calculated to identify abnormally high or low correlations. Based on the correlation score matrix, the hierarchical clustering algorithm was used to group the attributes. The silhouette coefficient was used to determine the optimal number of clusters, and the termination condition of the hierarchical clustering was controlled by setting the minimum cluster size to form attribute association clusters. Each association cluster represents a group of highly correlated cigarette attributes. The t-SNE algorithm was used to project high-dimensional attribute vectors onto a two-dimensional plane to generate a visualization of attribute relationships, which intuitively displayed the correlation strength and clustering results between different cigarette attributes. For the five consumer segments obtained by clustering, the maximum spanning tree algorithm was used to construct cigarette attribute association maps. The mutual information values ​​between the 12 cigarette attributes were calculated, and the optimal mutual information threshold was determined to be 0.3 through five-fold cross validation. Taking group A as an example, an attribute association map containing 9 nodes and 8 edges was obtained. The node2vec algorithm was applied for graph embedding, and the optimal hyperparameters were determined through grid search: the vector dimension was 64, the window size was 10, and the walk length was 80. A 64-dimensional attribute vector representation was obtained, such as the price sensitivity attribute vector [0.21,-0.15,...,0.33]. The cosine similarity between attribute vectors was calculated to obtain a 12×12 correlation score matrix. In the matrix, the correlation score between price sensitivity and purchase frequency was 0.82, indicating a high correlation. The mean of the correlation score was calculated to be 0.45, and the standard deviation was 0.18. Three pairs of attribute pairs with abnormally high correlation, such as scores > 0.81, and two pairs of attribute pairs with abnormally low correlation, such as scores < 0.09, were identified. The hierarchical clustering algorithm was used to group the attributes. The optimal number of clusters was determined to be 4 by the silhouette coefficient, and the minimum cluster size was set to 2. Four attribute association clusters were obtained, such as {price sensitivity, purchase frequency, income level} as one cluster. The t-SNE algorithm was used to project the 64-dimensional attribute vector onto a two-dimensional plane to generate a visualization diagram that intuitively presents the relative position and clustering results between the attributes. For example, price sensitivity and purchase frequency are close in the figure and are both located in the upper left corner.

[0026] In order to build a cigarette attribute association graph for consumer segment groups, the cigarette attributes of each segment group are defined, each attribute is used as a node in the graph, and the relationship between attributes is used as an edge. By analyzing the distance between attributes, the correlation of attribute combinations is identified. If the distance between two attributes is small, the correlation between the two attributes is strong.

[0027] The characteristics of consumer segment groups are obtained, and the cigarette attribute sets of each group are defined according to the characteristics; the association map of the cigarette attribute set is stored in the Neo4j graph database, and each attribute in the cigarette attribute set is used as a node in the graph, and the relationship between the attributes is used as an edge; the nodes in the attribute association map are mapped to a low-dimensional vector space using the Node2Vec algorithm to obtain the vector representation of each attribute node; the Euclidean distance between the attribute vectors is calculated according to the vector representation of the attribute node, and a distance matrix is ​​constructed; if the distance in the distance matrix is ​​less than a preset distance threshold, it is determined to be a strongly correlated attribute; a hierarchical clustering algorithm is used to group the strongly correlated attributes to form an attribute combination; the attribute association map is visualized by Gephi software to generate a visualization result of the attribute association map, in which the size of the node represents the importance of the attribute, and the thickness of the edge represents the strength of the relationship between the attributes.

[0028] Specifically, according to the characteristics of consumer segmentation groups, a specific set of cigarette attributes is defined for each group. The Neo4j graph database is used to store the attribute association map. Each attribute is used as a node in the graph, and the relationship between attributes is used as an edge. The weight of the edge is determined by calculating the Pearson correlation coefficient of the attribute value. The Node2Vec algorithm is used to map the nodes in the attribute association map to a low-dimensional vector space. The optimal vector dimension is determined by five-fold cross-validation to obtain the vector representation of each attribute node. The Euclidean distance between attribute vectors is calculated, and a distance matrix is ​​constructed. The interquartile range method is used to determine the distance threshold, and the distance less than the first quartile is determined as a strong correlation. Based on the distance matrix, a hierarchical clustering algorithm is used to group the attributes to form an attribute combination. The optimal number of clusters is determined by calculating the silhouette coefficient of different cluster numbers, and the minimum cluster size is set to 2 to control the termination condition of the hierarchical clustering, and the final attribute combination result is obtained. Each combination represents a group of cigarette attributes with strong correlation. The attribute association graph constructed was visualized using Gephi software to generate graph visualization results. The importance of the attribute was represented by the size of the node, and the strength of the relationship between the attributes was represented by the thickness of the edge. The relationship strength and clustering results of different cigarette attributes were intuitively displayed. For a certain consumer segment, 10 specific cigarette attributes were defined, including price sensitivity and taste preference, and the attribute association graph was stored using Neo4j graph database. The edge weight was determined by calculating the Pearson correlation coefficient of the attribute value. For example, the correlation coefficient between price sensitivity and purchase frequency was 0.75, indicating a strong correlation. The Node2Vec algorithm was applied for graph embedding. The optimal vector dimension was determined to be 64 through five-fold cross validation, and a 64-dimensional vector representation of each attribute node was obtained. The Euclidean distance between attribute vectors was calculated to construct a 10×10 distance matrix. The interquartile range method was used to determine the distance threshold. The first quartile was 0.3, and a distance less than 0.3 was considered a strong correlation. Based on the distance matrix, Ward's method was used for hierarchical clustering. By calculating the silhouette coefficients of different numbers of clusters, from 2 to 8, the optimal number of clusters was determined to be 4, with a silhouette coefficient of 0.68. The minimum cluster size was set to 2, and 4 attribute combinations were obtained, such as {price sensitivity, purchase frequency, income level} as a combination. The attribute association map was visualized using Gephi software. The node size was set according to the PageRank value, ranging from 10 to 50 pixels, and the edge thickness was set according to the correlation coefficient, ranging from 1 to 5 pixels. The visualization results show that the price sensitivity node is the largest, such as 50 pixels, and the connection edge with the purchase frequency is the thickest, such as 5 pixels, which intuitively reflects the relationship strength between the attributes and the clustering results.

[0029] Step S104, based on the cigarette attribute association map of each segmented group, the association pattern of cigarette attribute preferences of different segmented groups is found, and the association rules of tobacco leaf moisture content preference, cigarette price sensitivity and appearance preference style of the high nicotine preference group, and the association rules of smoke nicotine amount preference range, taste preference type and brand loyalty of the low tar preference group are obtained, so as to form a representative consumer portrait.

[0030] The April algorithm is used to extract high-frequency attribute combinations, and the optimal minimum support threshold is obtained through five-fold cross validation. The minimum support threshold is used to screen attribute association patterns with an occurrence frequency greater than a preset threshold. According to the high-frequency attribute combinations, the FP-Growth algorithm is used to analyze the association relationship between the attribute combinations, and the optimal minimum confidence and minimum lift thresholds are determined through grid search. The optimal minimum confidence and minimum lift thresholds are used to obtain strong association rules that meet the conditions. For the high-nicotine preference group and the low-tar preference group, the association rules of specific attribute combinations are extracted from the strong association rules, and the association rules of the specific attribute combinations are used to calculate the group association strength. The association rules are converted into group feature vectors, and one-hot encoding is used to obtain binary features, and the binary features are used to represent group attribute preferences. If the binary feature dimension is too high, principal component analysis is used to reduce the dimension of the binary feature to obtain a feature vector after dimension reduction. The K-means algorithm is used to cluster the feature vector after dimension reduction, and the optimal number of clusters is determined by calculating the silhouette coefficients of different numbers of clusters. The optimal number of clusters is used to form a representative consumer portrait.

[0031] Specifically, for the cigarette attribute association map of each segmented group, the April algorithm is used to extract high-frequency attribute combinations, and the optimal minimum support threshold is determined through five-fold cross-validation to screen out the attribute association patterns with high frequency. The FP-Growth algorithm is used to analyze the association relationship between high-frequency attribute combinations, and the optimal minimum confidence and minimum lift thresholds are determined through grid search to obtain strong association rules that meet the conditions. According to the association rules mined, the association rules of specific attribute combinations for the high nicotine preference group and the low tar preference group are extracted, and the association strength of the specific attribute combinations of the two groups is calculated using conditional probability. The association rules are converted into group feature vectors, and the rules are converted into binary features using one-hot encoding. Based on the group feature vectors, the principal component analysis is used to reduce the dimensionality of high-dimensional features, and the principal components with an explained variance ratio of more than 85% are retained. The K-means algorithm is used to cluster the reduced-dimensional features, and the optimal number of clusters is determined by calculating the silhouette coefficients of different numbers of clusters to form a representative consumer portrait. Use the radar chart visualization tool to generate a portrait radar chart, select the attribute with the largest contribution to the principal component as the dimension of the radar chart, and intuitively display the attribute preference characteristics of different groups.

[0032] Aiming at the attribute association map of a certain group of cigarette consumers, the April algorithm was used to extract high-frequency attribute combinations. The optimal minimum support threshold was determined to be 0.15 through five-fold cross validation, and 87 high-frequency attribute combinations were screened out. The FP-Growth algorithm was then used to analyze the association relationship between these combinations. The grid search obtained the optimal minimum confidence of 0.6 and the minimum lift of 1.2, and 53 strong association rules were obtained. For the group with high nicotine preference, the rule of "moisture content of tobacco leaves 12-14% → low price sensitivity of cigarettes" was extracted with a conditional probability of 0.78; the rule of "nicotine content of smoke 0.6-0.8mg → light taste preference" was found in the group with low tar preference, with an association strength of 0.85. These rules were converted into 200-dimensional binary feature vectors through one-hot encoding. Principal component analysis was used to reduce the dimension of the features, retaining the first 8 principal components, and the cumulative explained variance ratio reached 87%. The K-means algorithm was used to cluster the features after dimensionality reduction, and the silhouette coefficients of cluster numbers 2 to 10 were calculated. The optimal number of clusters was determined to be 5, and the silhouette coefficient was 0.68. Finally, the original attributes corresponding to the six principal components with the largest contributions, such as price sensitivity and nicotine preference, were selected as radar chart dimensions to generate consumer portrait radar charts for the five groups, which intuitively displayed the characteristic distribution of each group in these six dimensions.

[0033] According to the attribute preference data of cigarette price, moisture content of tobacco leaves, nicotine content of smoke, flavor type, and appearance style, the attribute data of different dimensions are normalized and mapped to the same scale space. The samples of the high nicotine preference group and the low tar preference group are resampled to balance the sample size of different sub-groups. By setting different support and confidence thresholds, the association pattern characteristics of different groups are adapted.

[0034] For cigarette attribute data, the minimum-maximum normalization method is used for normalization processing to obtain a standardized attribute data set. According to the standardized attribute data set, the SMOTE algorithm is used to resample the samples of the high nicotine preference group and the low tar preference group to generate a balanced sample data set. If the balanced sample data set is obtained, the April algorithm is applied to it for association rule mining, and the support and confidence threshold combinations are set by the grid search method. From the association rule mining results, the lift of each rule is calculated to determine whether it is greater than the preset threshold. If the rule lift is greater than the preset threshold, the rule is screened as a valid association pattern. For the screened effective association patterns, ten-fold cross validation is used to evaluate their stability and generalization ability, and determine the rules that perform stably on the validation set.

[0035] Specifically, the minimum-maximum normalization method is used to normalize attribute data such as cigarette price, tobacco leaf moisture content, smoke nicotine content, flavor type, and appearance style, and attribute data of different dimensions are mapped to a unified interval from 0 to 1 to obtain a standardized attribute data set. According to the standardized attribute data, the SMOTE algorithm is used to resample the samples of the high nicotine preference group and the low tar preference group, and the number of samples of the two groups is balanced by generating synthetic samples. The number of balanced samples is set to 1.5 times the number of samples of the larger group to obtain a balanced sample data set. The April algorithm is applied to the balanced sample data set for association rule mining, and different combinations of support and confidence thresholds are set by the grid search method. The support threshold range is set to 0.1 to 0.5, with a step size of 0.1; the confidence threshold range is set to 0.5 to 0.9, with a step size of 0.1. Association rule mining was performed for the high nicotine preference group and the low tar preference group respectively, and the number of rules, average confidence and other indicators of different groups were compared to adapt to the characteristics of association patterns of different groups. Based on the association rules obtained by mining, the lift of each rule was calculated, and the rules with a lift greater than 1 were screened out as valid association patterns. For possible contradictory rules, the rule confidence was used as the weight to merge the rules to form a characteristic association rule set for the high nicotine preference group and the low tar preference group. Ten-fold cross validation was used to evaluate the stability and generalization ability of the association rules, and the rules that performed stably on the validation set were retained. When processing cigarette attribute data, the minimum-maximum normalization method was first used to map the cigarette price from the 50-200 yuan range to the 0-1 range, the tobacco leaf moisture content from 10%-18% to 0-1, and the smoke nicotine content from 0.5-1.5mg to 0-1. Then, the SMOTE algorithm was used to resample the high nicotine preference group, with an original sample size of 5000, and the low tar preference group, with an original sample size of 3000, and the sample size of both groups was adjusted to 7500. The April algorithm was applied to the balanced samples for association rule mining, and the support threshold (0.1, 0.2, 0.3, 0.4, 0.5) and confidence threshold (0.5, 0.6, 0.7, 0.8, 0.9) were set by grid search, with a total of 25 combinations. The high nicotine preference group obtained the optimal rule set at a support of 0.2 and a confidence of 0.7, containing 87 rules and an average confidence of 0.82; the low tar preference group obtained the optimal rule set at a support of 0.3 and a confidence of 0.8, containing 62 rules and an average confidence of 0.89. The rule lift was calculated and the rules with lift greater than 1 were screened. 73 rules were retained for the high nicotine preference group and 58 rules were retained for the low tar preference group. For contradictory rules, such as "high price → high quality" with a confidence of 0.8 and "high price → low quality" with a confidence of 0.6, a weighted confidence combination was used to obtain "high price → high quality" with a weighted confidence of 0.71.Through ten-fold cross validation, the rules with confidence fluctuation less than 10% on the validation set were retained, and finally 65 rules were retained for the high nicotine preference group and 52 rules were retained for the low tar preference group. These rules constitute the feature association rule set of the two groups, reflecting the preference characteristics of different consumer groups.

[0036] Step S105, optimizing the combination of association rules based on the association rules of the high nicotine preference group and the low tar preference group, and obtaining a segmented group portrait model by balancing the relationship between different attributes and taking into account the optimal association combination of consumer preferences within the segmented groups.

[0037] The weighted average method is used to calculate the rule weight, in which the support, confidence and lift are respectively assigned weights of preset proportions to obtain a weighted association rule set. According to the weighted association rule set, an association rule network is constructed using the adjacency matrix representation method, with attributes as nodes, rules as edges, and the weights of edges set to the comprehensive weights of the corresponding rules. The Louvain algorithm is applied to the association rule network for community discovery. By adjusting the resolution parameter and the step size, the community division result with the largest modularity is selected to identify the closely related attribute clusters. If an attribute cluster is identified, one-hot encoding is used to convert each attribute cluster into a binary feature vector, and a multidimensional feature vector of the segmented group is constructed in combination with the original demographic characteristics. For the multidimensional feature vector, a sensitivity analysis method is used to determine the importance weight of each attribute in the portrait model by changing each attribute value one by one and observing the degree of influence on the final portrait. According to the sensitivity analysis results, different weights are assigned to each attribute in the feature vector to form a weighted feature vector. The silhouette coefficient is used to evaluate the clustering effect of the portrait model. By comparing the silhouette coefficients under different weight combinations, the optimal weight configuration is selected to obtain a segmented group portrait model.

[0038] Specifically, for the association rules of the high nicotine preference group and the low tar preference group, the weighted average method was used to calculate the rule weights, and the support, confidence and lift were assigned weights of 0.3, 0.4 and 0.3 respectively. The weight of each rule was calculated comprehensively to obtain the weighted association rule set. The adjacency matrix representation method was used to construct the association rule network, with attributes as nodes and rules as edges. The weight of the edge was set to the comprehensive weight of the corresponding rule to generate a rule network graph. The Louvain algorithm was applied to the constructed rule network graph for community discovery. By adjusting the resolution parameter from 0.5 to 2.0 with a step size of 0.1, the community division result with the largest modularity was selected to identify closely related attribute clusters, each of which represented a group of interrelated cigarette attribute preference characteristics. Based on the identified attribute clusters, one-hot encoding was used to convert each attribute cluster into a binary feature vector, and combined with the original demographic characteristics, a multidimensional feature vector of the segmented group was constructed. The sensitivity analysis method was used to determine the importance weight of each attribute in the portrait model by changing each attribute value one by one and observing the degree of influence on the final portrait. According to the results of sensitivity analysis, different weights are assigned to each attribute in the feature vector to form a weighted feature vector. The silhouette coefficient is used to evaluate the clustering effect of the portrait model. By comparing the silhouette coefficients under different weight combinations, the optimal weight configuration is selected to finally form a segmented group portrait model.

[0039] The weighted average calculation was performed on the 100 association rules of the high nicotine content preference group and the 80 association rules of the low tar content preference group. For example, the rule "high nicotine content → strong taste" has a support of 0.3, a confidence of 0.8, and a lift of 1.5, and the weighted comprehensive weight is 0.87. A 10×10 adjacency matrix is ​​used to represent the rule network, where the matrix element a[i][j] represents the rule weight from attribute i to attribute j, such as a[high nicotine content][strong taste] = 0.87. The Louvain algorithm was used for community discovery, and the resolution parameter was set from 0.5 to 2.0, with a step size of 0.1, for a total of 16 iterations, and the maximum modularity was obtained at a resolution of 1.3.

[0040] 0.68, and 5 attribute clusters were identified. The attribute clusters were converted into 25-dimensional binary feature vectors, such as cluster 1 [1,0,1,1,0], which means that the cluster contains attributes 1, 3, and 4. Combining 5 demographic characteristics such as age and income, a 30-dimensional feature vector was constructed. A sensitivity analysis was performed. For example, if the preference for nicotine content increased by 10%, the portrait similarity changed by 5%, and a weight of 0.5 was assigned accordingly; if the price sensitivity increased by 10%, the similarity changed by 3%, and a weight of 0.3 was assigned accordingly. These weights were used to weight the 30-dimensional feature vectors to obtain weighted feature vectors. The silhouette coefficients under different weight combinations were calculated, and the optimal combination obtained a silhouette coefficient of 0.72, based on which the final segmented group portrait model was determined.

[0041] Step S106, applying the segmented group portrait model to actual cigarette product development and marketing decisions, predicting the preferences of the high nicotine preference group and the low tar preference group for new product attribute combinations through the portrait model, and formulating product formula adjustment and marketing promotion strategies based on the demographic attribute distribution of different preference groups.

[0042] Obtain the importance weight of each attribute of the cigarette product, the importance weight is determined by the attribute importance in the portrait model; construct a cigarette product attribute scoring function based on the importance weight, the scoring function uses a weighted summation method to comprehensively score each attribute; input the attribute values ​​of each new product into the scoring function to obtain the preference score of the target group for the new product; use the preference score as the dependent variable and the product attribute as the independent variable to construct a multivariate linear regression model; if the variance inflation factor is greater than a preset threshold, regularize the corresponding variable; determine the influence of different attributes on the preference score based on the regression coefficient of the multivariate linear regression model; optimize the product formula using a grid search method, the grid search method takes maximizing the target group preference score as the optimization goal to obtain the optimal product attribute combination; obtain the demographic attribute distribution of the target group, use the optimal product attribute combination and the target group characteristics as input, and use the decision tree algorithm to construct a marketing strategy decision model; determine targeted marketing promotion strategy recommendations based on the output results of the decision tree algorithm; perform sensitivity analysis on the marketing promotion strategy recommendations to determine the influence of different attribute changes on the final decision.

[0043] Specifically, according to the segmented group portrait model, a cigarette product attribute scoring function is constructed, and the scores of each attribute are synthesized by weighted summation. The weights are determined according to the importance of the attributes in the portrait model. The numerical values ​​of each attribute of the new product are input into the scoring function to obtain the preference scores of the high nicotine preference group and the low tar preference group for the new product. Using the preference score as the dependent variable and the product attribute as the independent variable, a multivariate linear regression model is constructed. During the construction process, the variance inflation factor is used to detect multicollinearity, and the variables with VIF greater than 10 are regularized. The degree of influence of different attributes on the preference score is analyzed by regression coefficients to identify key influencing factors. Based on the results of the regression model, the grid search method is used to optimize the product formula, and the attribute value range is set as a constraint condition. The optimization goal is to maximize the preference score of the target group and obtain the optimal product attribute combination. The stability of the model is evaluated using five-fold cross validation, and the model with the smallest average prediction error is selected. Combined with the distribution of demographic attributes of different preference groups, a marketing strategy decision model is constructed using the decision tree algorithm. The optimal product attribute combination and target group characteristics are used as input to output targeted marketing promotion strategy recommendations. Set decision rules to convert decision tree results into specific marketing strategies, such as "if the age range is 25-35 years old, use social media promotion." Conduct sensitivity analysis to evaluate the impact of different attribute changes on the final decision and identify key decision factors.

[0044] When constructing the scoring function of cigarette product attributes, the weights of the four attributes of nicotine content, tar content, moisture content, and price were set to 0.3, 0.25, 0.25, and 0.2, respectively. The new products of the high nicotine content preference group were scored, and the preference score was 0.85. Using these scores to construct a multivariate linear regression model, the variance inflation factor was calculated and it was found that the VIF values ​​of tar content and nicotine content were 12 and 15, respectively, exceeding the threshold of 10. Ridge regression was applied to these two variables for regularization. The regression analysis results showed that the coefficient of influence of nicotine content on the preference score was the largest, which was 0.6. The grid search method was used to optimize the product formula, setting the nicotine content range to 0.8-1.2mg, tar content to 6-10mg, moisture content to 12-16%, and price to 50-100 yuan, with step sizes of 0.1mg, 1mg, 1%, and 10 yuan, respectively, to obtain the optimal combination: nicotine content 1.1mg, tar content 8mg, moisture content 14%, and price 80 yuan. The model was evaluated using a five-fold cross validation, with an average prediction error of 0.03. A decision tree marketing strategy model was constructed, with the optimal product attributes and target group characteristics such as age of 25-35 years old and monthly income of 8,000-12,000 yuan input, and the output strategy recommendation was "using social media promotion to emphasize the taste characteristics of the product." Sensitivity analysis showed that a 1% change in nicotine content resulted in a 5% probability of a decision change, which was the most influential factor.

[0045] The above embodiment is only one of the preferred implementation modes of the present invention and should not be used to limit the protection scope of the present invention. Any changes or modifications that are made to the main design concept and spirit of the present invention and have no substantive significance, and the technical problems they solve are still consistent with the present invention, should be included in the protection scope of the present invention.

Claims

1. A method for analyzing cigarette consumers, characterized in that: The method comprises: Obtain massive amounts of cigarette consumption data, pre-process the data, and obtain a structured consumption data set through data cleaning and data integration operations. The consumption data set includes cigarette price sensitivity, cigarette flavor preference type, cigarette appearance preference style, cigarette brand loyalty, and cigarette purchase frequency. The cigarette flavor preference type includes the preferred value of tobacco leaf moisture content, the preferred range of smoke nicotine content, and the preferred interval attributes of smoke tar content; According to the cigarette attributes in the consumption data set, consumers are clustered and grouped, and consumers with similar attribute preferences are divided into the same segment group. The cigarette attributes include cigarette price sensitivity, cigarette flavor preference type, cigarette appearance preference style, cigarette brand loyalty, and cigarette purchase frequency; For each consumer segment group obtained by clustering, a cigarette attribute association map of the group is constructed respectively, and the low-dimensional vector representation of the cigarette attributes within each segment group in the map is learned. The correlation between different attributes is measured by the distance between vectors. Based on the association maps of cigarette attributes of each segmented group, the association patterns of cigarette attribute preferences of different segmented groups were found, and the association rules of tobacco leaf moisture content preference, cigarette price sensitivity and appearance preference style of the high nicotine content preference group, and the association rules of smoke nicotine content preference range, flavor preference type and brand loyalty of the low tar content preference group were obtained, forming a representative consumer portrait; According to the association rules of the high nicotine preference group and the low tar preference group, the association rules are optimized and combined. By balancing the relationship between different attributes and taking into account the optimal association combination of consumer preferences within the segmented group, a segmented group portrait model is obtained. The segmented group portrait model is applied to the actual cigarette product research and development and marketing decision-making. The portrait model is used to predict the preferences of the high-nicotine preference group and the low-tar preference group for new product attribute combinations, and combined with the demographic attribute distribution of different preference groups, product formula adjustment and marketing promotion strategies are formulated.

2. The method according to claim 1, characterized in that The massive cigarette consumption data is obtained, the data is preprocessed, and a structured consumption data set is obtained through data cleaning and data integration operations. The consumption data set includes cigarette price sensitivity, cigarette flavor preference type, cigarette appearance preference style, cigarette brand loyalty, and cigarette purchase frequency, wherein the cigarette flavor preference type includes the tobacco leaf moisture content preference value, the smoke nicotine amount preference range, and the smoke tar amount preference interval attributes, including: Obtain consumers’ price sensitivity, brand loyalty, and purchase frequency data, use the K-means clustering algorithm to group consumers, and obtain the taste preference feature vectors of different consumer groups; Extracting visual features such as color and texture from the cigarette appearance preference style data according to the taste preference feature vector, and classifying the appearance style using a support vector machine; If the support vector machine classification is completed, a content-based recommendation algorithm is used to generate a personalized appearance style and taste combination recommendation list for each consumer group; Obtain geographic location distribution data and seasonal consumption trends, and use ARIMA time series analysis to predict cigarette sales in each region; Based on the sales forecast results and promotional activity responsiveness data, a random forest model is constructed to quantify the nonlinear impact of promotional activities on sales, and the optimal promotion strategy combination is determined by feature importance; The optimal promotion strategy combination is applied to the personalized recommendation list to obtain the final cigarette product sales plan.

3. The method according to claim 1, characterized in that According to the cigarette attributes in the consumption data set, consumers are clustered and grouped, and consumers with similar attribute preferences are divided into the same segmented group. The cigarette attributes include cigarette price sensitivity, cigarette flavor preference type, cigarette appearance preference style, cigarette brand loyalty, and cigarette purchase frequency, including: Perform standardization processing on each attribute data in the consumption data set to obtain standardized attribute data; The principal component analysis method is used to reduce the dimension of the standardized attribute data and obtain the feature vector after dimension reduction; Using the silhouette coefficient method to analyze the eigenvector after dimension reduction to determine the optimal number of clusters; According to the optimal number of clusters, K-means cluster analysis is performed on the feature vector after dimension reduction to obtain segmented groups; For the segmented groups, the average value and standard deviation of each attribute in the group are calculated to construct a group characteristic description matrix; Based on the group feature description matrix, a consumer classification model is constructed using a decision tree algorithm to determine the segmented group to which the newly added consumer belongs; If there are missing values ​​in the newly added consumer data, multiple imputation methods are used to handle the missing values; The newly added consumer data is input into the consumer classification model, and the segment group to which the newly added consumer belongs is determined by comparing the attribute value with the threshold of the decision node.

4. The method according to claim 1, characterized in that: For each consumer segment group obtained by clustering, a cigarette attribute association map of the group is constructed respectively, and a low-dimensional vector representation of the cigarette attributes within each segment group in the map is learned, and the correlation between different attributes is measured by the distance between vectors, including: The maximum spanning tree algorithm is used to construct the cigarette attribute association map, and the connection relationship between attribute nodes is determined by calculating the mutual information value between attributes. Applying the node2vec algorithm to perform graph embedding according to the cigarette attribute association graph to obtain a low-dimensional vector representation of the attribute node; The distance between attribute vectors is calculated using cosine similarity, and the distance value is normalized to between 0 and 1 to obtain the correlation score matrix between different attributes; If the correlation score matrix has been obtained, a hierarchical clustering algorithm is used to group the attributes to obtain attribute association clusters; For the attribute association cluster, the t-SNE algorithm is used to project the high-dimensional attribute vector onto a two-dimensional plane to generate an attribute relationship visualization diagram, which is used to display the correlation strength and clustering results between different cigarette attributes; It also includes: constructing a cigarette attribute association map for consumer segment groups, defining the cigarette attributes for each segment group, using each attribute as a node in the map, and the relationship between attributes as an edge. By analyzing the distance between attributes, the correlation of attribute combinations is identified. If the distance between two attributes is small, the correlation between the two attributes is strong.

5. The method according to claim 4, characterized in that The cigarette attribute association map for constructing consumer segment groups is described as follows: for each segment group, its cigarette attribute is defined, each attribute is used as a node in the graph, and the relationship between the attributes is used as an edge. By analyzing the distance between the attributes, the correlation of the attribute combination is identified. If the distance between the two attributes is small, the correlation between the two attributes is strong, including: Obtain the characteristics of consumer segment groups and define the cigarette attribute set for each group based on the characteristics; A Neo4j graph database is used to store the association graph of the cigarette attribute set, each attribute in the cigarette attribute set is used as a node in the graph, and the relationship between the attributes is used as an edge; Use the Node2Vec algorithm to map the nodes in the attribute association graph to a low-dimensional vector space to obtain the vector representation of each attribute node; Calculate the Euclidean distance between attribute vectors according to the vector representation of the attribute nodes, and construct a distance matrix; If the distance in the distance matrix is ​​less than a preset distance threshold, it is determined to be a strongly correlated attribute; Using a hierarchical clustering algorithm to group the strongly correlated attributes to form attribute combinations; The attribute association map is visualized by using Gephi software to generate a visualization result of the attribute association map, in which the size of the node represents the importance of the attribute, and the thickness of the edge represents the strength of the relationship between the attributes.

6. The method according to claim 1, characterized in that Based on the cigarette attribute association map of each segmented group, the association pattern of cigarette attribute preferences of different segmented groups is found, and the association rules of tobacco leaf moisture content preference, cigarette price sensitivity and appearance preference style of the high nicotine content preference group, and the association rules of smoke nicotine content preference range, taste preference type and brand loyalty of the low tar content preference group are obtained, forming a representative consumer portrait, including: The Apriori algorithm is used to extract high-frequency attribute combinations, and the optimal minimum support threshold is obtained through five-fold cross validation. The minimum support threshold is used to filter attribute association patterns whose occurrence frequency is greater than a preset threshold. According to the high-frequency attribute combination, the FP-Growth algorithm is used to analyze the correlation between the attribute combinations, and the optimal minimum confidence and minimum lift thresholds are determined through grid search. The optimal minimum confidence and minimum lift thresholds are used to obtain strong association rules that meet the conditions; For the high nicotine content preference group and the low tar content preference group, extracting association rules of specific attribute combinations from the strong association rules, and the association rules of the specific attribute combinations are used to calculate group association strength; The association rules are converted into group feature vectors, and one-hot encoding is used to obtain binary features, where the binary features are used to represent group attribute preferences; If the binary feature dimension is too high, principal component analysis is used to reduce the binary feature dimension to obtain the reduced feature vector; Clustering the reduced-dimensional feature vectors using a K-means algorithm, and determining an optimal number of clusters by calculating silhouette coefficients of different numbers of clusters, wherein the optimal number of clusters is used to form a representative consumer portrait; It also includes: normalizing attribute data of different dimensions according to cigarette price, tobacco moisture content, smoke nicotine content, flavor type, and appearance style attribute preference data, mapping the attribute data to the same scale space, resampling samples of high nicotine preference groups and low tar preference groups, balancing the sample size of different sub-groups, and adapting to the association pattern characteristics of different groups by setting different support and confidence thresholds.

7. The method according to claim 6, characterized in that According to the attribute preference data of cigarette price, tobacco leaf moisture content, smoke nicotine content, flavor type, and appearance style, attribute data of different dimensions are normalized, attribute data are mapped to the same scale space, samples of high nicotine content preference group and low tar content preference group are resampled, sample sizes of different subdivided groups are balanced, and different support and confidence thresholds are set to adapt to the characteristics of association patterns of different groups, including: For cigarette attribute data, the minimum-maximum normalization method is used to normalize them and obtain the standardized attribute data set; According to the standardized attribute data set, the SMOTE algorithm is used to resample the samples of the high nicotine preference group and the low tar preference group to generate a balanced sample data set; If the balanced sample data set is obtained, the Apriori algorithm is applied to it to mine association rules, and the support and confidence threshold combinations are set by the grid search method; From the association rule mining results, calculate the improvement of each rule and determine whether it is greater than the preset threshold; If the rule improvement is greater than the preset threshold, the rule is filtered as a valid association mode; For the screened effective association patterns, ten-fold cross validation is used to evaluate their stability and generalization ability, and to determine the rules that are stable on the validation set.

8. The method according to claim 1, characterized in that The association rules of the high nicotine preference group and the low tar preference group are optimized and combined, and the segmented group portrait model is obtained by balancing the relationship between different attributes and taking into account the optimal association combination of consumer preferences within the segmented group, including: The weighted average method is used to calculate the rule weights, where the support, confidence and lift are assigned preset weights respectively to obtain the weighted association rule set; According to the weighted association rule set, an association rule network is constructed using an adjacency matrix representation method, with attributes as nodes, rules as edges, and edge weights set to the comprehensive weights of corresponding rules; Applying the Louvain algorithm to the association rule network for community discovery, by adjusting the resolution parameter and the step size, selecting the community division result with the largest modularity, and identifying closely related attribute clusters; If attribute clusters are identified, one-hot encoding is used to convert each attribute cluster into a binary feature vector, and combined with the original demographic characteristics, a multidimensional feature vector of the segmented group is constructed; For the multidimensional feature vector, a sensitivity analysis method is used to determine the importance weight of each attribute in the portrait model by changing each attribute value one by one and observing the degree of influence on the final portrait; According to the sensitivity analysis results, different weights are assigned to each attribute in the feature vector to form a weighted feature vector; The silhouette coefficient is used to evaluate the clustering effect of the portrait model. By comparing the silhouette coefficients under different weight combinations, the optimal weight configuration is selected to obtain the segmented group portrait model.

9. The method according to claim 1, characterized in that: The segmented group portrait model is applied to actual cigarette product development and marketing decisions. The portrait model is used to predict the preferences of the high nicotine preference group and the low tar preference group for new product attribute combinations, and combined with the demographic attribute distribution of different preference groups, product formula adjustment and marketing promotion strategies are formulated, including: Obtaining the importance weight of each attribute of the cigarette product, wherein the importance weight is determined by the attribute importance in the portrait model; Constructing a cigarette product attribute scoring function according to the importance weights, wherein the scoring function uses a weighted summation method to comprehensively calculate the scores of each attribute; Inputting the attribute values ​​of the new product into the scoring function to obtain the preference score of the target group for the new product; Using the preference scores as dependent variables and product attributes as independent variables, a multiple linear regression model is constructed; If the variance inflation factor is greater than the preset threshold, the corresponding variable is regularized; Determining the degree of influence of different attributes on the preference score based on the regression coefficients of the multivariate linear regression model; A grid search method is used to optimize the product formula, wherein the grid search method takes maximizing the preference score of the target group as the optimization goal to obtain the optimal product attribute combination; Obtain the demographic attribute distribution of the target group, take the optimal product attribute combination and target group characteristics as input, and use the decision tree algorithm to build a marketing strategy decision model; Determine targeted marketing promotion strategy recommendations based on the output results of the decision tree algorithm; A sensitivity analysis is recommended for the marketing promotion strategy to determine the impact of changes in different attributes on the final decision.

Images