A tobacco commercial enterprise benchmarking grouping method based on capacity constraint K-means clustering and feature parameter optimization
By using capacity-constrained K-means clustering and feature parameter optimization, combined with cross-validation of multiple supervised learning algorithms, the problems of uncontrollable capacity, weak model interpretability, and insufficient robustness of results in the grouping of tobacco commercial enterprises were solved, achieving scientific, fair, and interpretable grouping results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FUJIAN TOBACCO CO LONGYAN CO
- Filing Date
- 2026-02-05
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies make it difficult to achieve controllable grouping capacity, interpretable models, and robust results in tobacco commercial enterprises. This results in grouping results that fail to meet the rigid requirements of management for grouping balance, and the grouping results are difficult to explain using business logic, forming a 'black box model' and reducing management credibility.
We employ a capacity-constrained K-means clustering algorithm, combined with feature parameter optimization and cross-validation of multiple supervised learning algorithms. Through data preprocessing, principal component contribution rate analysis, and cluster quality index evaluation, we ensure the balance and interpretability of the grouping results.
This approach achieves a balance between the scientific validity and fairness of the grouping results, enhances the robustness and interpretability of the results, and allows grouping differences to be directly attributed to core business indicators, breaking the 'black box model' and improving management credibility.
Smart Images
Figure CN122132933A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary field of enterprise management and data science, and in particular to a method for benchmarking and grouping tobacco commercial enterprises based on capacity-constrained K-means clustering and feature parameter optimization. Background Technology
[0002] Benchmarking is a core tool for improving corporate performance, and its effectiveness rests on "comparability among similar entities." County-level bureaus under tobacco commercial enterprises exhibit inherent heterogeneity in terms of geographical environment, economic foundation, and business structure, making traditional administrative classification or simple clustering methods insufficient to meet the needs of refined benchmarking.
[0003] Existing technologies face three main challenges: (1) Uncontrollable grouping capacity: Classic clustering algorithms (such as K-means) aim at the optimal distance, resulting in "giant clusters" and "micro clusters" in the grouping results, which cannot meet the rigid requirements of management for grouping balance. Subsequent manual adjustments will undermine the scientific and objective nature of the algorithm.
[0004] (2) Weak model interpretability: In pursuit of comprehensiveness, a large number of indicators are invested, which can easily lead to the "curse of dimensionality" and overfitting, making it difficult to explain the grouping results with business logic, forming a "black box model" and reducing management credibility.
[0005] (3) Insufficient robustness of results: Existing methods lack a mechanism for external and multi-angle verification of the robustness of clustering results, making it difficult to ensure that the grouping pattern is a true reflection of the intrinsic structure of the data, rather than an accidental product of the algorithm.
[0006] Therefore, there is an urgent need for a benchmarking grouping technology that can deeply integrate management logic with statistical optimization and ensure that the results are scientific, fair, and verifiable. Summary of the Invention
[0007] In view of this, the purpose of this invention is to provide a benchmarking grouping method for tobacco commercial enterprises based on capacity-constrained K-means clustering and feature parameter optimization, and to provide a benchmarking grouping method that integrates constraint clustering, feature optimization and intelligent verification, aiming to systematically solve the above challenges.
[0008] To achieve the above objectives, the present invention adopts the following technical solution: a method for benchmarking and grouping tobacco commercial enterprises based on capacity-constrained K-means clustering and feature parameter optimization, comprising the following steps: Step 1: Data preprocessing and sample screening. County-level tobacco commercial enterprises are classified into primary categories according to business type, and a homogeneous sample set is screened out. Step 2: Use the capacity-constrained K-means clustering algorithm to cluster the samples under the constraints of the preset number of groups and the volume of each group; Step 3: Based on principal component contribution rate analysis, select the core parameter set from multi-dimensional feature parameters; Step 4: Evaluate the grouping effect using the clustering quality index and compare the performance of the full-parameter model and the optimized parameter model; Step 5: Cross-validate the clustering results using multiple supervised learning algorithms to evaluate the robustness and interpretability of the grouping.
[0009] In a preferred embodiment, the business types mentioned in step 1 include county-level bureaus that sell only tobacco products and county-level bureaus that sell tobacco products. The data comes from the provincial company's unified financial, business and monopoly management system and has undergone missing value processing, outlier cleaning and Z-score standardization preprocessing.
[0010] In a preferred embodiment, the objective function of the capacity-constrained K-means clustering algorithm in step 2 is:
[0011] The constraint condition is that the upper limit of the sample size of each group is satisfied. Hard capacity control is achieved through iterative allocation and centroid updates.
[0012] In a preferred embodiment, the specific process for selecting the feature parameters in step 3 includes: performing principal component analysis on the full parameter dataset to calculate the contribution rate of each original parameter to the overall variance; selecting parameters whose contribution rate is significantly higher than the average level to construct a core parameter set.
[0013] In a preferred embodiment, the core parameter set includes per capita cigarette sales volume D1, per capita total tax and profit E1, and per capita cigarette sales revenue E3, with its contribution rate ranking among the top three out of 24 parameters. In a preferred embodiment, the clustering quality index Q in step 4 is calculated using the following formula:
[0014] In the formula, the total variance (TSS) measures the difference between all samples and the overall mean, the within-group variance (WSS) measures the difference between samples within a group and the mean of that group, the between-group variance (BSS) measures the difference between the means of each group and the overall mean, and Q>0.8 indicates excellent grouping performance.
[0015] In a preferred embodiment, the supervised learning algorithm in step 5 includes Support Vector Machine (SVM), Random Forest (RF), Backpropagation Neural Network (BPNN), and Naive Bayes (NB), and its classification accuracy is used to verify the robustness of the clustering results.
[0016] The present invention also provides a computer device, characterized in that it comprises: The processor, memory, and bus, wherein the memory stores machine-readable instructions executed by the processor; When the computer device is running, the processor communicates with the memory via a bus, and when the machine-readable instructions are executed by the processor, they perform a tobacco commercial enterprise benchmarking grouping method based on capacity-constrained K-means clustering and feature parameter optimization as described above.
[0017] The present invention also provides a computer-readable storage medium, characterized in that it comprises: The computer-readable storage medium contains a computer program; The computer program is executed by the processor as described in the method for benchmarking and grouping tobacco commercial enterprises based on capacity-constrained K-means clustering and feature parameter optimization.
[0018] Compared with the prior art, the present invention has the following beneficial effects: Scientific rigor and fairness are combined: the capacity constraint algorithm ensures the balance of grouping, and feature optimization improves the business interpretability of grouping, so that the benchmarking results are both mathematically optimal and meet management fairness.
[0019] Strong robustness: Through cross-validation using multiple machine learning methods, it was independently confirmed from the outside that the grouping pattern is a robust reflection of the intrinsic structure of the data, rather than random noise.
[0020] High interpretability: Group differences can be directly attributed to a few core business indicators, breaking the "black box model" and making the results easy for managers and business units to understand and accept.
[0021] Good universality: This methodological framework is not only applicable to the benchmarking grouping of tobacco commercial enterprises, but can also be transferred to other classification scenarios with similar balance needs, such as the division of tobacco ecological zones and the classification of retail customers. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of PCA projection visualization analysis of the pure sales area grouping model Y1 in a preferred embodiment of the present invention; Figure 2 The contribution rate of Y1 in the pure sales area grouping model of the preferred embodiment of the present invention (taking the top 20 parameters of contribution rate); Figure 3 PCA projection visualization analysis of the Y2 clustering effect of the pure sales area grouping model in a preferred embodiment of the present invention; Figure 4 This is a parallel coordinate diagram of the Y2 grouping parameters of the pure sales area grouping model (standardized values after dimensionless elimination) of a preferred embodiment of the present invention. Figure 5A comparison of the grouping parameters (original values) of the pure sales area grouping model Y2 in a preferred embodiment of the present invention. Figure 6 The confusion matrix (n=51) of the fully parameter-based machine learning model is a preferred embodiment of the present invention. Figure 7 The confusion matrix of the machine learning model based on preferred parameters in a preferred embodiment of the present invention (n=51); Figure 8 This is a schematic diagram of the innovative method system formed by the present invention, which is a preferred embodiment of the present invention. Detailed Implementation
[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0024] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0025] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.
[0026] A method for benchmarking and grouping tobacco commercial enterprises based on capacity-constrained K-means clustering and feature parameter optimization, referenced Figure 1-8 This invention adopts a design science research paradigm, with the core objective of constructing and validating an integrated methodology for solving the benchmarking grouping problem of county-level tobacco commercial bureaus. The entire research design is based on empirical data from 51 pure sales district county-level bureaus in Province F, strictly following the logical chain of "problem definition - scheme design - evaluation and verification," and is described below in two parts: methods and results and analysis.
[0027] Part One: Method, including the following steps: Step 1: Research Sample and Data Sources To ensure the homogeneity of the groupings, this invention first classifies all county-level bureaus under the F Province tobacco commercial system based on their core business structure. Specifically, they are divided into: 51 "pure sales county-level bureaus" (whose main business is solely cigarette marketing) and 25 "dual-tobacco county-level bureaus" (which simultaneously engage in tobacco production and cigarette marketing). This pre-classification based on the essence of the business effectively eliminates incomparable differences caused by fundamentally different business structures, allowing subsequent cluster analysis to be conducted within relatively homogeneous subgroups, consistent with the statistical principle of stratified sampling.
[0028] This invention focuses on the group of county-level sales bureaus (n=51), and systematically collects years of basic data (Table 1) on 24 indicators across six dimensions reflecting their overall operations. All raw data comes from the provincial company's unified financial, business, and monopoly management systems, ensuring data authority and consistency. Before analysis, this invention rigorously cleans the data (handling missing and outliers) and preprocesses it using Z-score standardization to eliminate the dominant influence of different indicator dimensions and orders of magnitude on distance calculations.
[0029] Table 1. Characteristic parameters of county-level bureaus (branches) in the F province tobacco commercial system that sell only tobacco products.
[0030] Step 2: Design of Capacity-Constrained K-means Clustering Algorithm To meet the rigid management requirement of F Province's benchmarking management work—namely, "precisely dividing 51 units into 6 benchmarking groups, with each group strictly controlled to have 8 to 9 units"—this invention fundamentally modifies the classic K-means algorithm, designing a capacity-constrained K-means algorithm. Its core idea is to add a capacity check and adjustment mechanism during the sample allocation phase of the iteration process. The specific formula is as follows:
[0031] st
[0032] For all i ≠ j For all i = 1, 2, ..., k For all i = 1, 2, ..., k Where, μ i Let μ represent the centroid of the i-th cluster, and μ represent the upper limit of capacity.
[0033] The specific algorithm steps are as follows: I. Input: Data point set X, number of groups K, and upper limit of capacity of each group {C1, C2, ..., Ck}.
[0034] II. Initialization: Select K initial centroids randomly or based on a specific strategy.
[0035] III. Iterate until convergence: Allocation steps (including capacity constraints): Calculate the distance from each data point to each centroid, and modify the traditional "nearest neighbor allocation" into a constrained optimized allocation process. First, sort all points by their distance to their nearest centroid; then, attempt to assign each point to its nearest centroid in turn. If the cluster is full (reaching its capacity limit), reassign the point to its "second nearest" centroid that is not yet full. This process continues until all points are allocated and the capacity constraints of each cluster are satisfied.
[0036] Update steps: Based on the allocation results, recalculate the centroid of each cluster (the mean of all points in the cluster).
[0037] IV. Output: K groups that satisfy the capacity constraint and their centroids.
[0038] The advantage of this algorithm lies in the fact that it does not obtain a mathematically "optimal" but managerially "infeasible" cluster, but rather treats the managerial constraints as boundary conditions of the optimization problem itself. Therefore, the final output is both managerially executable (strictly satisfying capacity balance) and statistically optimal under these constraints, truly achieving a deep integration of managerial logic and statistical optimization.
[0039] Step 3: Feature parameter optimization design based on contribution rate To avoid the "curse of dimensionality" and noise interference caused by high-dimensional data, and to improve the interpretability of the model, it is necessary to select the core feature set driving the grouping from the initial multiple feature parameters (24 parameters in this invention). This invention abandons the selection method that relies on subjective experience and designs a data-driven, reproducible feature parameter optimization process.
[0040] (1) Contribution Rate-Based Analysis: Principal Component Analysis (PCA) is performed on the standardized full-parameter dataset. PCA can transform the original correlated variables into a series of uncorrelated linear combinations (principal components) and reveal the contribution of each original variable to the variance of these principal components. This invention calculates the overall contribution rate of each original parameter to the principal components. This contribution rate quantifies the amount of information carried by the parameter in characterizing the main differences between samples. Parameters with high contribution rates are key characteristics that distinguish different groups of county-level bureaus.
[0041] (2) Feature parameter optimization design: Based on contribution rate ranking, this invention designs two sets of competitive clustering models for comparative experiments: Baseline model (Y1): Input consists of all 24 parameters. This model represents a "full information" model without considering feature parameter selection.
[0042] Optimization Model (Y2): The input is a set of core feature parameters selected based on contribution rate analysis. The selection criteria are: parameters whose contribution rate is significantly higher than the average level and whose business implications are complementary.
[0043] By comparing the clustering results of Y1 and Y2, this invention can scientifically evaluate the necessity of feature selection and determine the optimal and concise grouping parameter system in a data-driven manner.
[0044] Step 4: Clustering Quality Assessment Criteria To objectively and quantitatively compare the performance of clustering models under different parameter combinations, this invention uses a clustering quality index based on variance decomposition as the core evaluation indicator. This index has clear statistical significance, and its calculation is based on the total sum of squares (TSS), within-cluster sum of squares (WSS), and between-cluster sum of squares (BSS), satisfying TSS = WSS + BSS. The clustering quality index Q is defined as: (1) In the formula, the total variance (TSS) measures the difference between all samples and the overall mean, the within-group variance (WSS) measures the difference between the samples within a group and the mean of that group, and the between-group variance (BSS) measures the difference between the means of each group and the overall mean.
[0045] The Q value ranges from [0,1]. The closer the Q value is to 1, the greater the proportion of between-group differences (BSS) relative to total differences (TSS), and the smaller the proportion of within-group differences (WSS), indicating a more ideal grouping effect. In practical applications, Q > 0.6 is generally considered to indicate a good clustering effect, while Q > 0.8 indicates an excellent effect with very clear separation between groups.
[0046] Step 5: Cross-validation of machine learning algorithms To ensure that the clustering results are not accidental and that the grouping patterns have a robust and learnable mapping relationship in the feature space, this invention introduces a supervised learning algorithm for cross-validation, forming a methodological closed loop of "unsupervised discovery - supervised verification".
[0047] The specific steps are as follows: The optimal grouping results obtained from the aforementioned capacity-constrained clustering (i.e., the output labels of model Y2) are considered as a set of "pseudo-true labels." Using this labeled data as the training set, four classic supervised learning classification models are constructed: (1) Support Vector Machine (SVM): The basic principle of SVM is to find an optimal hyperplane that maximizes the separation of samples of different classes in the feature space. For nonlinear problems, mapping is performed through a kernel function, as shown in equation (2).
[0048] (2) In the formula, φ(X) is the mapping function. Since the mapping function has a complex form, it is difficult to calculate its inner product. Therefore, the kernel method can be used, that is, the inner product of the mapping functions is defined as the kernel function: This avoids the explicit calculation of the inner product.
[0049] This invention is implemented using the fitcsvm function in the Statistics and Machine Learning Toolbox of MATLAB, selects the radial basis function (RBF) as the kernel function, and uses grid search and cross-validation to optimize the hyperparameter C (penalty coefficient) and kernel scale.
[0050] (2) Random Forest Regression Model Construction (Random Forest, RF): RF is an ensemble learning algorithm that makes predictions by constructing multiple decision trees and combining their outputs (majority vote or average). Each tree uses bootstrap sampling to extract samples during training, and selects the optimal split point from a randomly selected feature subset when splitting nodes, in order to enhance the diversity and generalization ability of the model, as shown in Equation (3).
[0051] y^=mode{h1(x),h2(x),...,h m (x)} (3) In the formula, y^ represents the prediction result of the random forest; h m (x) represents the predictions of m decision trees for sample X; M is the total number of decision trees; mode is the selection of the majority of trees.
[0052] This invention is implemented using MATLAB's TreeBagger function, with the number of trees set to 100 and the other parameters using default values.
[0053] (3) Back Propagation Neural Network (BPNN) Model Construction: BPNN is a multi-layer feedforward neural network that adjusts weights through an error backpropagation algorithm. This invention uses MATLAB's Neural Network Toolbox to construct a BPNN with one hidden layer. The number of neurons in the hidden layer is determined according to the method of Zhang Pei et al., and the Levenberg-Marquardt training algorithm is used. The model's convergence error is set to 0.0001, the learning rate to 0.05, and the maximum number of training iterations to 1000.
[0054] (4) Naive Bayes model (NB): The NB classifier is based on Bayes' theorem and makes predictions under the assumption of feature conditional independence, as shown in equation (4).
[0055] (4) In the formula, This is the posterior probability, which is the probability that a sample belongs to a class similar to C after observing feature X. Likelihood is the probability that a feature will appear when a sample belongs to class C. This is the prior probability, i.e., the probability distribution of category C in the population; Evidence factor, which is the probability of feature X appearing in all categories.
[0056] This invention is implemented using the MATLAB function fitcnb, following the method of Zhang Pei et al., and the characteristic distribution is assumed to be a Gaussian distribution.
[0057] The above algorithms are used to learn the mapping relationship between "feature parameters and group labels," and then the original data is regrouped for prediction. The predicted results are compared with the original clustering groups. To comprehensively evaluate the performance of the classification model, this invention uses a confusion matrix and model accuracy for evaluation. The confusion matrix is a C×C square matrix (C is the number of categories), where the columns represent the true categories and the rows represent the predicted categories. Accuracy is defined as the proportion of correctly predicted samples to the total number of samples. High accuracy means that the supervised machine learning model can reproduce the grouping structure of unsupervised clustering from the selected feature parameters, thus providing strong external evidence that the grouping pattern is not random noise, but an inherent and robust topological structure in the feature space. In-depth analysis of misclassified samples can further reveal cases at the group boundaries or with special business models, providing additional insights for management.
[0058] Part Two: Results and Analysis S1: Clustering performance evaluation and problem diagnosis of the baseline model (Y1) This invention evaluated the baseline model Y1, which uses all 24 parameters as input. This model represents the grouping performance under unoptimized "full information" input. As shown in Table 2, the clustering quality index Q of model Y1 is only 0.4271. After variance decomposition, it was found that the within-group sum of squares (WSS=658.85) of model Y1 is significantly greater than the between-group sum of squares (BSS=491.15), forming a serious "inverted" classification phenomenon between within-group and between-group. This data indicates that the differences between units within the six groups formed after clustering with 24 feature parameters as input exceed the average differences between different groups, and the grouping model fails to effectively capture and separate the systematic differences between samples. This signifies that the clustering performance of model Y1 is a failure and cannot meet the grouping requirement of "homogeneous within groups and heterogeneous between groups".
[0059] Table 2. Evaluation of Clustering Effect of Grouping Model for County-Level Bureaus (Branch Companies) in Pure Sales Districts (n=51)
[0060] Note: The closer Q is to 1, the greater the difference between groups and the smaller the difference within groups, indicating a better grouping effect. Generally speaking, ≥0.6 is considered a good clustering effect.
[0061] To further understand the grouping results of model Y1, this invention uses principal component projection plots ( Figure 1 Visual analysis revealed that, except for the relatively clustered sample points in group 5, the sample points in the other five groups exhibited extensive overlap and interweaving in two-dimensional space, lacking clear boundaries between groups. This indicates that under full parameter input, a large number of irrelevant or redundant feature parameters generate severe noise interference, leading to distortion of the group distance metric. Consequently, the clustering algorithm fails to discover meaningful data structures, and its output becomes uninterpretable. This result provides strong counter-evidence to the notion that "more data = better model," demonstrating the critical importance of paying close attention to the risks associated with the "curse of dimensionality" and "overfitting" during data analysis.
[0062] S2: Feature Contribution Rate Analysis and Core Parameter Identification To find the optimal path for the model, this invention decomposed the contribution rates of all 24 parameters involved in clustering based on principal component analysis. Results ( Figure 2The results show that different parameters explain significant differences in sample variance. The contribution rates exhibit a "long-tail distribution," meaning that the vast majority of effective grouping information is concentrated on a few core parameters. The three parameters with the highest contribution rates are, in order: D1 (per capita cigarette sales volume), E1 (per capita total tax and profit), and E3 (per capita cigarette sales revenue). These three parameters directly point to the core and essential operational efficiency and value creation capabilities of county-level bureaus in pure sales areas: per capita sales volume reflects the scale effect of consumer satisfaction; per capita tax and profit is the final financial outcome; and per capita sales revenue combines sales volume and structure, reflecting market capacity. These three characteristic parameters have extremely clear business implications and are strongly correlated with each other (sales volume drives sales revenue, which in turn affects tax and profit), together forming a "golden triangle" for measuring the operational performance of units in pure sales areas.
[0063] This finding illustrates that in benchmarking and grouping scenarios, not all available indicators have equal classification power. Those "driving indicators" directly related to core management objectives and reflecting the essential attributes of the organization have a much higher information density than "descriptive indicators" that describe the background environment or process details. Feature selection through contribution rate analysis is essentially a data-driven process of "feature denoising" and "business focus."
[0064] S3: Construction and Interpretation of the Optimization Model (Y2) Based on the above findings, this invention further constructs an optimization model Y2, whose input consists only of three selected core parameters: D1, E1, and E3. Applying the same capacity-constrained K-means clustering algorithm (K=6, group size 8-9), this invention obtains entirely new grouping results.
[0065] The performance evaluation (Table 2) shows that the optimized model Y2 achieved a significant performance improvement. Its clustering quality index Q jumped from 0.4271 in model Y1 to 0.9535, an increase of 123.3%. Variance decomposition data shows that the within-group sum of squares (WSS) decreased to 6.98, while the between-group sum of squares (BSS) remained relatively high at 143.02. This data comparison leads to a clear conclusion: when using a simplified core parameter set, within-group heterogeneity is compressed to a very low level, while between-group differences are fully amplified. Further analysis of the principal component projection plots of the grouping results for model Y2 (Table 2) is also provided. Figure 3 As can be seen, the six groups are precisely segmented with clear edges. Sample points in each group are densely distributed around their centroids, while significant distances are maintained between the centroids of different groups. There is almost no overlap between groups. This clear spatial separation intuitively demonstrates the superiority of the optimized model's grouping performance.
[0066] Combined with parallel coordinate graph ( Figure 4 Comparison chart with grouping parameters () Figure 5This invention further interprets the business meaning of the grouping results. Parallel coordinate graph ( Figure 4 The chart shows that the six lines representing the six groups exhibit a highly regular, stepped arrangement on the three coordinate axes D1, E1, and E3. The high-performing groups (such as Group 6) are at the highest level in all three indicators, while the low-performing groups (such as Group 3) are at the lowest level, with the intermediate groups progressing sequentially. More importantly, these lines remain almost parallel across the three dimensions, with very few intersections. This indicates that for a county-level bureau, its performance on these three core performance indicators is highly coordinated and rises and falls synchronously. (Group Comparison Chart) Figure 5 This directly displays the original mean values of each group, further confirming the continuity of differences between the characteristic parameter groups. This perfect alignment between mathematical representation and business logic greatly enhances the interpretability and management credibility of the grouping results. Managers can clearly understand that the differences in grouping are essentially differences in operational performance levels, and the grouping results directly map the relative position of each unit in the industry value chain.
[0067] S4: Machine Learning Cross-Validation Analysis To independently evaluate the robustness of the optimal grouping result (model Y2) from an external perspective, this invention employs supervised machine learning for cross-validation. This invention constructs feature parameter sets based on both full parameters (24 parameters) and optimized parameters (3 parameters), and trains four classification models—SVM, RF, BPNN, and NB—using the grouping labels of model Y2 as the training objective. The results are as follows: Figure 6 , Figure 7 As shown.
[0068] (1) The confusion matrix of the full-parameter machine learning model shows that the prediction results of each model have a low matching degree with the true labels (the results of model Y2), and the accuracy is generally poor. This confirms that in high-dimensional noisy data, supervised machine learning models also have difficulty finding stable classification rules, which echoes the low clustering quality of model Y1.
[0069] (2) The confusion matrix of the machine learning models based on the optimized parameters shows that the prediction results of each model have been significantly improved, especially the BPNN model, which has improved its accuracy by 28.94%. At the same time, the RF model and the BPNN model have both achieved a decision accuracy of 96.08%. The above results indicate that these two types of models have basically learned and reproduced the mapping rules between the three core parameters and the six group labels. The predictions for all 51 samples are basically consistent with the original clustering results, which further proves that there is a definite, stable decision boundary between the low-dimensional feature space composed of D1, E1, and E3 and the final grouping, which can be captured by the machine learning algorithm. The clustering results are not a product of random algorithm, but a true reflection of the inherent structure of the data.
[0070] (3) Misjudgment Analysis and Management Implications of Other Models: The accuracy rates of the SVM and NB models were 64.71% and 94.12%, respectively. In-depth analysis of their misjudged samples not only did not compromise the reliability of the models but also generated additional management value. For example, the feature combinations of individual units misjudged by the Bayesian model showed that their ratio of "sales per capita" to "tax and profit per capita" deviated from the mainstream pattern (e.g., moderate sales but extremely high tax and profit, or vice versa), which may indicate their unique market structure, cost control capabilities, or contributions from non-cigarette businesses. Units misjudged by the SVM were mostly located near the group boundaries, and their feature values were almost identical to the centroid distance between their assigned group and neighboring groups. This revealed the existence of "marginal cases" on the statistical boundaries under hard capacity constraints. These findings suggest that managers may not be able to adequately benchmark within groups for these units and require more personalized analysis and attention.
[0071] In summary, machine learning cross-validation, from an external mathematical perspective, confirms that the grouping results obtained by combining capacity-constrained clustering with feature parameter optimization possess extremely high robustness, consistency, and learnability. The reproducibility of results from the BPNN and RF models constitutes solid evidence of the method's effectiveness, while the misjudgment analysis of other models forms a data-driven management diagnostic, transforming simple statistical grouping into a diagnostic tool that supports refined management.
[0072] In summary, this invention establishes a complete, reproducible, and verifiable benchmarking grouping methodology. Empirical verification has successfully solved a core challenge in benchmarking grouping for county-level tobacco bureaus: how to scientifically, fairly, and interpretably group heterogeneous county-level bureaus while satisfying the rigid management constraint of "grouping capacity balance." By introducing a capacity-constrained K-means algorithm, feature parameter optimization based on contribution rate, and machine learning cross-validation, this invention constructs a complete methodological system and has achieved significant results on data from county-level bureaus in pure sales areas of Province F. The core findings can be summarized in three points: First, full parameter input leads to the failure of model Y1 (Q=0.4271), revealing the harm of high-dimensional noise; second, focusing on core business parameters (sales volume per employee, tax and profit per employee, and sales revenue per employee) can significantly improve model performance (Q=0.9535), thereby promoting high homogeneity within groups and clear separation between groups; third, the grouping results can be cross-validated by external models, with all four supervised machine learning models demonstrating excellent reproducibility (up to 96.08%), proving that the grouping results are not statistically random, but a robust reflection of the inherent structure of the data. Most importantly, the subtle differences in reproducibility accuracy among different algorithms and their misjudgment analysis translate into deeper management insights.
[0073] 1. Capacity-constrained clustering algorithm: a dialectical unity of management practicality and statistical scientificity.
[0074] The dilemma of applying traditional clustering in management grouping essentially lies in the conflict between its unconstrained optimization objective and management requirements. The capacity-constrained K-means algorithm designed in this invention is innovative not because it proposes a completely new mathematical problem, but because it creatively introduces the concept of "capacity constraint" from operations research into the typical scenario of benchmarking grouping, providing an efficient and feasible heuristic solution. The algorithm's advantages are: through pre-designed capacity constraints, it ensures the scientific rigor, consistency, and learnability of the clustering process, ultimately achieving balanced group sizes and minimizing the distance between units within a group under constraints. Its management implications lie in making scientific methods serve management purposes, enabling benchmarking grouping to form an automated scientific computing process based on management rules, significantly improving the efficiency and objectivity of grouping decisions. This solves the long-standing fundamental problem of difficulty in organizing benchmarking activities and unfair evaluations due to large disparities in group size.
[0075] 2. Feature optimization mechanism: Achieving synergy between data dimensionality reduction and business interpretability.
[0076] The significant difference in accuracy between models Y1 and Y2 is one of the core empirical contributions of this invention. Feature parameter selection based on principal component contribution rates is essentially a supervised dimensionality reduction method (although using unsupervised PCA as a tool, the goal is to improve supervised clustering results). It successfully compresses 24 complex and ambiguous feature parameters into 3 concise and clear business parameters. This brings two benefits: mathematically, it eliminates noise, reduces the "curse of dimensionality" in distance metrics, and allows the inherent structure of the data (performance hierarchy) to clearly emerge; business-wise, it forms a highly interpretable grouping logic. Grouping differences can be directly and unambiguously attributed to three core performance indicators—"per capita cigarette sales," "per capita total tax and profit," and "per capita cigarette sales revenue"—all of which can be clearly interpreted and understood. This breaks down the trust barriers of "machine illusion" and "black box models," enabling the grouping results to be accepted, discussed, and applied to subsequent management actions by business departments and county-level bureaus. This interpretability is the management influence generated by the statistical model, ensuring that benchmarking management is no longer anchored to the core key feature parameters driving high-quality development.
[0077] 3. Machine Learning Cross-Validation: Closed-loop Enhancement and Deep Mining of Unsupervised and Supervised Learning.
[0078] Traditional clustering evaluation relies heavily on internal metrics such as silhouette coefficient and Davies-Bouldin index. While these metrics have some reference value, they cannot answer the fundamental question of whether the grouping pattern is genuine and robust. This invention introduces four supervised learning algorithms—Support Vector Machine (SVM), Random Forest (RNF), Neural Network (NN), and Naive Bayes (NB)—for cross-validation, constructing a closed-loop data analysis system of "grouping-validation-mining." Its importance is reflected in three aspects: First, it provides a set of external, objective, and multifaceted evaluation criteria. Model accuracy is a more intuitive and less manipulable metric than internal metrics. The high accuracy of the RF and BPNN models (96.08%), and the equally excellent performance of the NB model (94.14%), strongly support the validity of the clustering results. This multiple validation greatly enhances the persuasiveness of the grouping results. Second, it reveals the broad learnability of grouping patterns within the mathematical field. High accuracy means that not only can the current data be accurately classified, but the classification rule can also be successfully learned and reproduced by a variety of algorithms with different principles (from tree-based ensemble learning to connection learning simulating neural networks, and Bayesian methods based on statistical probability). This indicates that the structure discovered in this invention is not specific to a particular dataset or algorithm, but rather a universal, robust analysis result based on the inherent data structure of the samples. Finally, and most importantly for management, the differences in misjudgments by different algorithms constitute data-driven management diagnostics. In-depth analysis of misjudged samples shows that misjudgments by SVM and NB are mostly concentrated in the group boundary regions in the feature space. The feature values of these "marginal cases" are very close to the centroid distances of their respective groups and neighboring groups, and their grouping itself has a certain degree of statistical ambiguity. Certain specific misjudgments may reveal nonlinear relationships between indicators or unique localized business models. This analysis process transforms simple model performance evaluation into a management diagnostic tool for identifying "management concerns" and "personalized analysis objects," achieving a deep spillover of evaluation value.
[0079] 4. Comparative advantages of optimized parameter models: a leap from "chaotic fitting" to "essential laws".
[0080] By comparing the validation results of machine learning models based on full parameters and those based on optimized parameters, this invention clearly reveals a watershed: the former, in a noisy, multi-dimensional feature parameter system, even advanced machine learning algorithms struggle to find stable patterns, resulting in low accuracy—essentially a manifestation of "chaotic fitting"; while the latter, in a key feature parameter system composed of core indicators, shows performance improvements across all algorithms, exhibiting stable, efficient, and consistent regularity. This conclusion indicates that the management level of an excellent management and control system lies not in the sheer number of monitoring indicators, but in its ability to accurately identify and focus on a few key levers driving organizational value. The three core feature parameters selected in this invention—"sales per capita, tax and profit per capita, and sales per capita"—constitute the "golden triangle" for evaluating the operational efficiency of county-level bureaus in pure sales districts. Based on this grouping, the operational management levels of each unit can be presented in a tiered manner most clearly and powerfully, thereby guiding the focus of benchmarking management from complex process details to the ultimate value creation capability and core operational performance. This not only enhances the scientific nature of the grouping but also strengthens the guiding role of benchmarking management as a "command stick" for strategic execution.
[0081] Deepening benchmarking management is a key approach for the tobacco industry to promote high-quality development and achieve efficient governance. This invention addresses three major challenges in benchmarking groups of county-level tobacco commercial bureaus: rigid management constraints, weak model interpretability, and insufficient robustness of results. It successfully constructs and validates a methodology that integrates "capacity-constrained K-means clustering," "contribution rate screening," and "machine learning cross-validation." Figure 8 ) The innovation of this invention is mainly reflected in three aspects: First, the problem definition is more precise, clearly defining the benchmarking grouping problem as "clustering with hard capacity constraints", which is more in line with the real-world scenario of benchmarking management in county-level bureaus of tobacco commercial enterprises; second, the solution is systematically integrated, not applying a single algorithm in isolation, but organically linking algorithm modification (constraint), feature parameter optimization (interpretability), and multi-model cross-validation to form a systematic computational chain that is interconnected and self-verifying; third, the value-oriented practical shift, always taking "management interpretability, decision supportability, and insight mining" as the highest principle of method design, rather than simply pursuing mathematical optimality, ensuring that the research results can be seamlessly embedded and empower actual benchmarking management.
[0082] The theoretical contribution of this invention lies in its systematic construction of a bridge connecting cutting-edge statistical machine learning methods with complex real-world management scenarios. It demonstrates how the creative integration of methodologies (constrained clustering, interpretable feature optimization, multi-algorithm cross-validation, and deep mining) can solve practical problems encountered by traditional methods in application. It provides a complete, transparent, operable, reproducible, and insightful scientific methodology that significantly improves the scientific rigor, fairness, and efficiency of grouping (classification) work, and directly transforms the analysis results into actionable management guidelines. The method developed in this invention is highly transferable; any field facing a similar problem of "scientifically and fairly classifying objects based on multidimensional features under a preset number and size of groups" can draw upon this method. For example, it can also be applied to specific scenarios such as tobacco leaf suitable ecological zone classification, rational distribution of retail customers, market order assessment, and precise profiling and grading of retail customers.
Claims
1. A method for benchmarking and grouping tobacco commercial enterprises based on capacity-constrained K-means clustering and feature parameter optimization, characterized in that, Includes the following steps: Step 1: Data preprocessing and sample screening. County-level tobacco commercial enterprises are classified into primary categories according to business type, and a homogeneous sample set is screened out. Step 2: Use the capacity-constrained K-means clustering algorithm to cluster the samples under the constraints of the preset number of groups and the volume of each group; Step 3: Based on principal component contribution rate analysis, select the core parameter set from multi-dimensional feature parameters; Step 4: Evaluate the grouping effect using the clustering quality index and compare the performance of the full-parameter model and the optimized parameter model; Step 5: Cross-validate the clustering results using multiple supervised learning algorithms to evaluate the robustness and interpretability of the grouping.
2. The method for benchmarking and grouping tobacco commercial enterprises based on capacity-constrained K-means clustering and feature parameter optimization according to claim 1, characterized in that, The business types mentioned in step 1 include county-level bureaus that sell only tobacco products and county-level bureaus that sell tobacco products. The data comes from the provincial company's unified financial, business and monopoly management system and has undergone missing value processing, outlier cleaning and Z-score standardization preprocessing.
3. The method for benchmarking and grouping tobacco commercial enterprises based on capacity-constrained K-means clustering and feature parameter optimization according to claim 1, characterized in that, The objective function of the capacity-constrained K-means clustering algorithm described in step 2 is: The constraint is that the upper limit of the sample size of each group is satisfied. Hard capacity control is achieved through iterative allocation and centroid updates.
4. The method for benchmarking and grouping tobacco commercial enterprises based on capacity-constrained K-means clustering and feature parameter optimization according to claim 1, characterized in that, The specific process for feature parameter selection in step 3 includes: performing principal component analysis on the full parameter dataset to calculate the contribution rate of each original parameter to the overall variance; selecting parameters whose contribution rate is significantly higher than the average level to construct a core parameter set.
5. The method for benchmarking and grouping tobacco commercial enterprises based on capacity-constrained K-means clustering and feature parameter optimization according to claim 4, characterized in that, The core parameter set includes per capita cigarette sales volume (D1), per capita total tax and profit (E1), and per capita cigarette sales revenue (E3), and its contribution rate ranks among the top three of the 24 parameters.
6. The method for benchmarking and grouping tobacco commercial enterprises based on capacity-constrained K-means clustering and feature parameter optimization according to claim 1, characterized in that, The formula for calculating the cluster quality index Q in step 4 is as follows: In the formula, the total variance (TSS) measures the difference between all samples and the total mean, the within-group variance (WSS) measures the difference between samples within a group and the mean of that group, the between-group variance (BSS) measures the difference between the means of each group and the total mean, and Q>0.8 indicates excellent grouping performance.
7. The method for benchmarking and grouping tobacco commercial enterprises based on capacity-constrained K-means clustering and feature parameter optimization according to claim 1, characterized in that, The supervised learning algorithms mentioned in step 5 include Support Vector Machine (SVM), Random Forest (RF), Backpropagation Neural Network (BPNN), and Naive Bayes (NB). Their classification accuracy is used to verify the robustness of the clustering results.
8. A computer device, characterized in that, include: The processor, memory, and bus, wherein the memory stores machine-readable instructions executed by the processor; When the computer device is running, the processor communicates with the memory via a bus, and when the machine-readable instructions are executed by the processor, they perform a tobacco commercial enterprise benchmarking grouping method based on capacity-constrained K-means clustering and feature parameter optimization as described in any one of claims 1 to 7.
9. A computer-readable storage medium, characterized in that, include: The computer-readable storage medium contains a computer program; The computer program is executed by the processor as described in any one of claims 1 to 7, which is a method for benchmarking and grouping tobacco commercial enterprises based on capacity-constrained K-means clustering and feature parameter optimization.