Multi-level supply chain business collaboration and distribution method and system

Abstract

This invention discloses a multi-level supply chain collaborative distribution method, comprising constructing a consumer demand-based collaborative cigarette distribution recommendation system framework, collecting and preprocessing multi-source review data, performing in-depth analysis of consumer demand characteristics based on the preprocessed data, clustering the analyzed demands and calculating product similarity, and adjusting the product distribution ratio based on the clustered demands and calculated product similarity. This invention, through multi-source data collection and in-depth analysis, accurately captures consumer demand characteristics, and combines cluster analysis and product similarity calculation to achieve dynamic and intelligent adjustment of product distribution ratios. This effectively improves the accuracy and flexibility of collaborative distribution, reduces supply-demand imbalances, optimizes supply chain resource allocation, enhances market responsiveness, and provides scientific methodological support for efficient collaboration in multi-level supply chains.

Description

Technical Field

[0001] This invention relates to the field of tobacco sales, and more specifically to a multi-level supply chain collaborative distribution method and system. Background Technology

[0002] Currently, China's tobacco industry primarily employs various cigarette distribution models, including distribution by grade level, distribution by price segment, distribution by grade level + price segment, selected distribution points, "quantitative labeling," and "customer demand." These traditional distribution methods are all based on in-depth data mining within the tobacco industry, relying heavily on the operational capabilities and historical purchase data of tobacco retailers. Similarly, collaborative decision-making between industrial and commercial enterprises has long depended on static distribution strategies driven by historical sales data. This model forms regional distribution plans by statistically analyzing monthly sales and inventory turnover rates for each retailer's various product specifications. However, with the increasing diversification and personalization of the consumer market, this method has revealed significant drawbacks: on the one hand, sales data only reflects the results of "transactions that have occurred," making it difficult to capture the dynamic changes in consumers' implicit needs such as price sensitivity and taste preferences; on the other hand, the micro-level demand differences in different regional markets are masked by the crude administrative hierarchy, resulting in the inability to meet the differentiated ordering needs of convenience stores, catering outlets, and entertainment venues within the same county-level unit. To address these needs, academia and industry have conducted numerous explorations. At the demand perception level, patent CN113807092A proposes a method for analyzing online reviews of cigarette brands based on the LDA topic model. By constructing a topic dictionary containing 150 keywords such as "tar content" and "throat irritation," it models user reviews on e-commerce platforms to extract demand features such as price sensitivity and packaging satisfaction. In the market segmentation field, patent CN107103488B designs a hybrid judgment model that integrates collaborative filtering and K-means clustering. Its innovation lies in introducing user profile features, transforming discrete variables such as age, occupation, and income level into coordinate points in a continuous vector space, and identifying similar consumer groups by calculating Euclidean distance. In addition, the paper "Research on the Precision Distribution Model of High-end and High-priced Cigarettes Based on Business Districts—Taking Yulin City as an Example" attempts to introduce GIS (Geographic Information System) technology, incorporating 12 indicators such as the resident population density, per capita GDP, and number of competing stores within a 3-kilometer radius of retailers into the business district evaluation system. Through on-site surveys, it delineates 6 customer tiers and formulates differentiated limited distribution strategies accordingly.

[0003] However, the existing cigarette distribution technologies mentioned above have significant limitations in terms of market demand response and precise decision-making. First, traditional static distribution strategies based on price ranges and other metrics rely excessively on explicit indicators such as historical sales data and inventory turnover rates. Essentially, this is a passive control mechanism based on "post-event attribution," making it difficult to effectively capture the dynamic evolution of consumers' implicit needs. For example, key consumer preferences such as price sensitivity and throat irritation can only be indirectly inferred from transaction data. This leads to the differentiated needs of heterogeneous scenarios such as convenience stores and restaurants within the same administrative region being masked by administrative hierarchical divisions, resulting in a crude management model of "supply-driven demand."

[0004] Demand perception technology suffers from a single-dimensional limitation: While LDA topic model-based review analysis can extract explicit topics such as price and packaging, it lacks quantitative modeling of professional dimensions such as taste intensity and tar compatibility, and does not consider the impact of review time decay factors on demand trend prediction; while collaborative filtering models that integrate user profiles are limited by the accuracy bottleneck of continuous transformation of discrete variables, and are prone to semantic drift in the representation learning of qualitative features such as occupation and income.

[0005] Furthermore, market segmentation methods have not yet broken through the limitations of spatial granularity. Although the GIS-driven business district evaluation system introduces geographic information parameters such as population density and competitor distribution, it has not established a mapping relationship between consumer demand feature vectors and spatial coordinates. This makes it difficult to accurately match the dynamic demand spectrum of terminal types such as convenience stores and entertainment venues with the division of the six customer tiers.

[0006] Finally, product substitution relationship modeling suffers from methodological flaws: current similarity calculations based on Euclidean distance or cosine similarity primarily focus on single attributes such as price and tar content, failing to construct a product feature vector space that incorporates multimodal features such as aroma characteristics and packaging design. This results in substitution recommendation strategies exhibiting significant lag in responding to changes in sensory demands, such as "spicy aftertaste" and "lingering sweetness." These technological bottlenecks collectively lead to the industry's long-standing structural contradictions, including supply-demand mismatch and low inventory turnover efficiency. Summary of the Invention

[0007] In view of the shortcomings of existing technologies, the purpose of this invention is to provide a multi-level supply chain industrial and commercial collaborative distribution method and system for solving key technical problems in the cigarette business field, such as insufficient accuracy of supply and demand matching, lagging dynamic response to demand, deviation in product substitution relationship modeling, and extensive regional market segmentation.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a multi-level supply chain business-industry collaborative distribution method, characterized by comprising the following steps:

[0009] Step 1: Construct a framework for a collaborative cigarette distribution recommendation system based on consumer demand;

[0010] Step two: Collect and preprocess multi-source comment data;

[0011] Step 3: Conduct in-depth analysis of consumer demand characteristics based on the preprocessed data from Step 2;

[0012] Step four: Cluster the requirements obtained in step three and calculate product similarity.

[0013] Step 5: Adjust the product specification distribution ratio based on the demand completed in Step 4 and the calculated product similarity.

[0014] As a further improvement to the present invention, the specific steps for multi-source comment data collection and preprocessing in step two are as follows:

[0015] Step Two: 1. Obtain cigarette product review data for the study area and construct an original review corpus. The corpus contains cigarette product names, user IDs, product ratings, and review texts;

[0016] Step 22: Compare the cigarette product names in the corpus with the product specifications that retailers in the region can sell, and remove product review data that cannot be sold in this region;

[0017] Steps two and three involve denoising the comment text in the corpus, including filtering advertising text, removing duplicate comments, eliminating entries not related to the target brand, and removing general praise, emotional outbursts, or irrelevant comments to obtain a valid set of comments. ;

[0018] Step 24: Use a Chinese word segmentation toolkit to... Lexical analysis was performed, and the word segmentation results were corrected by combining the results with a tobacco industry terminology dictionary to generate a standardized comment summary set. ;

[0019] Step 25, to Perform entity recognition and construct a user-comment association table. The associated table includes user ID, comment time, comment summary, and comment product.

[0020] As a further improvement to the present invention, the specific steps for in-depth analysis of consumer demand characteristics in step three are as follows:

[0021] Step 31: Load the fine-tuned BERT model specifically for the tobacco industry. Each comment in the dataset undergoes fine-grained sentiment analysis, outputting a four-dimensional feature vector:

[0022]

[0023] Among them, PS represents price sensitivity, TS represents taste intensity, DS represents packaging satisfaction, and JS represents tar compatibility.

[0024] Step 32, to Apply attention-based weighted aggregation to calculate the performance of a single user. Demand feature vector :

[0025]

[0026] in, Indicates the first The time decay factor of a comment is calculated based on two dimensions: semantic importance and time decay, using a learnable parameter matrix. and bias To obtain semantic importance, the comment feature vectors are mapped to the attention score space:

[0027]

[0028] The tanh function is used for nonlinear activation, making the distribution more concentrated.

[0029] Introducing the calculation of time decay coefficient Time decay factor, because recent comments are more valuable than historical comments:

[0030]

[0031] Step 3: Compile statistics on the specifications of each cigarette brand. Average demand feature vector:

[0032]

[0033] in, This indicates that the regulations apply to cigarette products. The user ID that made the comment.

[0034] As a further improvement to the present invention, the specific steps for clustering and product similarity calculation in step four are as follows:

[0035] Step 41: Output the region clustering results Each cluster contains regional planning and demand characteristic distribution.

[0036] Step 42: Extract the list of retailer product specifications within the regional planning scope of this cluster, and for each product specification that the retailer can sell. Based on its characteristics, the degree to which each requirement is satisfied is quantified, resulting in... ,in, The higher it is, the more it meets the requirement k.

[0037] Step Four Three: Calculate the comprehensive demand score of each product to measure the overall matching degree of the product to the consumer demand:

[0038]

[0039] Among them, n represents the number of consumer demands, represents the weight of the th demand, which is obtained from the ratio of this demand to the total number of demands.

[0040] Step Four Four: Construct the product feature vector space:

[0041]

[0042] Use One - Hot encoding to process discrete variables;

[0043] Step Four Five: Calculate the similarity between the product specifications sold by each retailer :

[0044]

[0045] Among them, is the price weight factor, represents the cosine similarity between tar and flavor type.

[0046] As a further improvement of the present invention, the specific steps for adjusting the product specification delivery ratio in Step Five are as follows:

[0047] Step Five One: Based on the similar products of each product specification, calculate the overall competitiveness of the similar products and this product specification;

[0048]

[0049] Balance its own needs and the influence of similar products:

[0050]

[0051] Among them , The larger it is, the more it emphasizes its own attributes;

[0052] Step Five Two: Under ideal conditions, calculate the delivery ratio completely based on the matching of consumer demands:

[0053]

[0054] Step Five Three: Integrate historical data, control the adjustment intensity, and update the delivery ratio:

[0055]

[0056] in, This represents the current distribution ratio for each product specification;

[0057] Step 54: Set a minimum distribution ratio to avoid product supply disruptions:

[0058]

[0059] Set minimum delivery ratio

[0060] And perform normalization:

[0061]

[0062] Step 55: Push campaign adjustment suggestions to enterprises. The interface protocol fields are defined as follows:

[0063] {

[0064] "Recommendation Number": "RECOM_2024_Q1"

[0065] "Adjustment Suggestions": {"Nanjing Yuhuashi": -15%, "Nanjing Xuanhemen": +20%}

[0066] "Valid Date": "2024-04-01"

[0067] }

[0068] Another aspect of the present invention provides a system comprising:

[0069] Data acquisition module: used to capture multi-source heterogeneous comment data and perform preprocessing;

[0070] Demand Feature Mining Module: Employs deep learning models to extract multidimensional consumer demand features;

[0071] Demand clustering and similar product calculation module: integrates demand characteristics and product similarity calculation to achieve micro-market segmentation;

[0072] Product specification allocation ratio adjustment module: Generates alternative combination suggestions based on product attributes and demand matching.

[0073] Compared with existing solutions, the present invention has the following advantages:

[0074] 1. Dynamic Demand Perception: Compared with the limitations of CN113807092A, which only uses the LDA topic model to extract explicit topics such as price and packaging, this invention innovatively adopts a fine-tuned BERT model specifically for the tobacco industry, combined with attention mechanism and time decay factor, to achieve fine-grained quantitative analysis of four-dimensional demand characteristics: price sensitivity, taste intensity, packaging satisfaction and tar compatibility.

[0075] 2. Multimodal Product Similarity Modeling: Overcoming the shortcomings of the collaborative filtering model in CN107103488B, which is based solely on discrete variables of user profiles, this invention constructs a hybrid similarity calculation model that includes price weights and cosine similarity. By encoding discrete attributes, a multidimensional feature vector space covering the core attributes of cigarettes is formed, which improves the accuracy of product substitution relationship identification, especially when dealing with changes in sensory needs such as "spicy aftertaste" and "long-lasting sweetness".

[0076] 3. Demand-driven matching mechanism: This invention outputs clusters with differentiated demand characteristics through a demand clustering algorithm, and generates a product-demand matching degree evaluation matrix by combining the list of products sold by retailers, thereby achieving spatial granularity refinement from administrative level division to demand characteristic mapping.

[0077] 4. Dynamic Deployment Strategy: To address the supply-demand mismatch problem in the traditional tiered deployment model, this invention proposes a balance model that integrates its own demand score with the competitiveness of similar products. It also designs a progressive adjustment algorithm that includes historical data fusion coefficients, minimum deployment constraints, and normalization processing. This avoids market fluctuations caused by aggressive adjustments while ensuring the matching degree of supply and demand. Attached Figure Description

[0078] Figure 1 This is a flowchart illustrating the multi-level supply chain business-industry collaborative distribution method of the present invention. Detailed Implementation

[0079] The present invention will now be described in further detail with reference to the embodiments shown in the accompanying drawings.

[0080] Reference Figure 1 As shown in the figure, a multi-level supply chain business-industry collaborative distribution method in this embodiment mainly includes the following steps.

[0081] S1: Construct a framework for a collaborative industrial and commercial cigarette distribution recommendation system based on consumer demand.

[0082] The system includes the following functional modules:

[0083] Data acquisition module: used to capture multi-source heterogeneous comment data and perform preprocessing;

[0084] Demand Feature Mining Module: Employs deep learning models to extract multidimensional consumer demand features;

[0085] Demand clustering and similar product calculation module: integrates demand characteristics and product similarity calculation to achieve micro-market segmentation;

[0086] Product specification allocation ratio adjustment module: Generates alternative combination suggestions based on product attributes and demand matching.

[0087] S2. Multi-source comment data collection and preprocessing

[0088] S2.1 Obtain cigarette product review data for the study area and construct an original review corpus. The corpus contains cigarette product names, user IDs, product ratings, and review texts;

[0089] S2.2 Compare the cigarette product names in the corpus with the product specifications that retailers in the region can sell, and remove product review data that cannot be sold in this region;

[0090] S2.3 Perform noise reduction processing on the comment text in the corpus, including filtering advertising text, deleting duplicate comments, removing entries that are not related to the target brand, and removing general praise, emotional outbursts, or irrelevant comments, to obtain a valid set of comments. ;

[0091] S2.4, Use a Chinese word segmentation toolkit to... Lexical analysis was performed, and the word segmentation results were corrected by combining the tobacco industry terminology dictionary (including professional descriptions such as "spicy aftertaste" and "lingering sweetness"), and a standardized comment summary set was generated. ;

[0092] S2.5, to Perform entity recognition and construct a user-comment association table. The associated table includes user ID, comment time, comment summary, and comment product.

[0093] S3. In-depth analysis of consumer demand characteristics

[0094] S3.1 Load the fine-tuned BERT model specifically for the tobacco industry, and... Each comment in the dataset undergoes fine-grained sentiment analysis, outputting a four-dimensional feature vector:

[0095]

[0096] Among them, PS represents price sensitivity, TS represents taste intensity, DS represents packaging satisfaction, and JS represents tar compatibility.

[0097] S3.2, To Apply attention-based weighted aggregation to calculate the performance of a single user. Demand feature vector :

[0098]

[0099] in Indicates the first The time decay factor for each comment is calculated based on two dimensions: semantic importance and time decay. This is achieved through a learnable parameter matrix. and bias To obtain semantic importance, the comment feature vectors are mapped to the attention score space:

[0100]

[0101] The tanh function is used for nonlinear activation, which makes the distribution more concentrated.

[0102] Introducing the calculation of time decay coefficient Time decay factor, because recent comments are more valuable than historical comments:

[0103]

[0104] S3.3, Statistics on various cigarette brands Average demand feature vector:

[0105]

[0106] in, This indicates that the regulations apply to cigarette products. The user ID that made the comment.

[0107] S4. Demand Clustering and Product Similarity Calculation

[0108] S4.1 Output region clustering results Each cluster contains regional planning and demand characteristic distribution.

[0109] S4.2 Extract the list of retailer product specifications within the regional planning scope of this cluster, and for each product specification that the retailer can sell. Based on its characteristics, the degree to which each requirement is satisfied is quantified, resulting in... ,in The higher the value, the more the requirement k is met.

[0110] S4.3 Calculate the overall demand score for each product to measure the product's performance. Overall match with consumer needs:

[0111]

[0112] Where n represents the number of consumer demands, Indicates the first The weight of a demand is determined by its proportion to the total number of demands.

[0113] S4.4 Constructing the product feature vector space:

[0114]

[0115] One-Hot encoding is used to process discrete variables;

[0116] S4.5 Calculate the similarity between the product specifications sold by each retailer. :

[0117]

[0118] in As a price weighting factor, This indicates the cosine similarity between tar and aroma type.

[0119] S5, Product Specification Distribution Ratio Adjustment

[0120] S5.1. Based on similar products for each product specification, calculate the overall competitiveness of similar products with this product specification.

[0121]

[0122] Balancing one's own needs with the impact of similar products:

[0123]

[0124] in , The larger it is, the more it values ​​its own attributes.

[0125] S5.2 Ideally, the allocation ratio should be calculated entirely based on matching consumer demand:

[0126]

[0127] S5.3 Integrate historical data, control the adjustment intensity, and update the allocation ratio:

[0128]

[0129] in, This represents the current distribution ratio for each product specification.

[0130] S5.4 Set a minimum distribution ratio to avoid product supply disruptions:

[0131]

[0132] Set minimum delivery ratio

[0133] And perform normalization:

[0134]

[0135] S5.5. Push delivery adjustment suggestions to enterprises. The interface protocol fields are defined as follows:

[0136] {

[0137] "Recommendation Number": "RECOM_2024_Q1"

[0138] "Adjustment Suggestions": {"Nanjing Yuhuashi": -15%, "Nanjing Xuanhemen": +20%}

[0139] "Valid Date": "2024-04-01"

[0140] }

[0141] In summary, the multi-level supply chain business-industry collaborative distribution method of this embodiment includes:

[0142] 1. Preprocessing method for comment data: By constructing a tobacco industry terminology dictionary to correct the Chinese word segmentation results, combining entity recognition technology to generate a standardized comment summary set, and establishing a user-comment association table, the unstructured comment data can be accurately cleaned and semantically enhanced.

[0143] 2. Demand Feature Extraction Based on Fine-tuned BERT Model: Fine-grained sentiment analysis of reviews is performed using a tobacco-specific BERT model, outputting four-dimensional feature vectors of price sensitivity, taste intensity, packaging satisfaction, and tar compatibility. A dynamic demand feature vector generation method is constructed by integrating time decay factor and semantic importance through an attention mechanism.

[0144] 3. Construction of Multimodal Product Feature Vector Space and Hybrid Similarity Calculation: By comprehensively utilizing encoding to process discrete attributes and combining price weights and cosine similarity indices, a multidimensional feature vector space covering the core attributes of cigarettes is constructed, solving the problem of insufficient fusion of multimodal information such as aroma features and packaging design in traditional similarity calculation.

[0145] 4. Demand-driven dynamic clustering and matching degree evaluation: By using a demand clustering algorithm and combining it with the list of products sold by retailers, a product-demand matching degree evaluation matrix is ​​generated, which overcomes the coarse granularity of demand feature space caused by traditional administrative level division.

[0146] 5. Dynamic Allocation Adjustment: Based on a balance model of product comprehensive demand score and the overall competitiveness of similar products, a progressive adjustment algorithm is designed, which includes historical data fusion, minimum allocation constraint and normalization processing, to achieve synergistic optimization of supply and demand matching degree and market stability.

[0147] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A multi-level supply chain business-industry collaborative distribution method, characterized in that: Includes the following steps: Step 1: Construct a framework for a collaborative cigarette distribution recommendation system based on consumer demand; Step two: Collect and preprocess multi-source comment data; Step 3: Conduct in-depth analysis of consumer demand characteristics based on the preprocessed data from Step 2; Step four: Cluster the requirements obtained in step three and calculate product similarity. Step 5: Adjust the product specification allocation ratio based on the demand identified in Step 4 and the calculated product similarity. The specific steps for in-depth analysis of consumer demand characteristics in Step 3 are as follows: Step 31: Load the fine-tuned BERT model specifically for the tobacco industry. Each comment in the dataset undergoes fine-grained sentiment analysis, outputting a four-dimensional feature vector: Among them, PS represents price sensitivity, TS represents taste intensity, DS represents packaging satisfaction, and JS represents tar compatibility. Step 32, to Apply attention-based weighted aggregation to calculate the performance of a single user. Demand feature vector : in, Indicates the first The time decay factor for each comment is calculated based on two dimensions: semantic importance and time decay; then, it is applied through a learnable parameter matrix. and bias Obtain semantic importance and map the comment feature vectors to the attention score space: The tanh function is used for nonlinear activation, making the distribution more concentrated; then, a time decay coefficient is introduced for calculation. Time decay factor; Step 33: Compile statistics on the specifications of each cigarette brand. Average demand feature vector: in, This indicates that the regulations apply to cigarette products. The user ID that made the comment.

2. The multi-level supply chain business-industry collaborative distribution method according to claim 1, characterized in that: The specific steps for multi-source comment data collection and preprocessing in step two are as follows: Step Two:

1. Obtain cigarette product review data for the study area and construct an original review corpus. The corpus contains cigarette product names, user IDs, product ratings, and review texts; Step 22: Compare the cigarette product names in the corpus with the product specifications that retailers in the region can sell, and remove product review data that cannot be sold in this region; Steps two and three involve denoising the comment text in the corpus, including filtering advertising text, removing duplicate comments, eliminating entries not related to the target brand, and removing general praise, emotional outbursts, or irrelevant comments to obtain a valid set of comments. ; Step 24: Use a Chinese word segmentation toolkit to... Lexical analysis was performed, and the word segmentation results were corrected by combining the results with a tobacco industry terminology dictionary to generate a standardized comment summary set. ; Step 25, to Perform entity recognition and construct a user-comment association table. The associated table includes user ID, comment time, comment summary, and comment product.

3. The multi-level supply chain business-industry collaborative distribution method according to claim 1 or 2, characterized in that: The specific steps for clustering and calculating product similarity in step four are as follows: Step 41: Output the region clustering results. Each cluster contains the distribution of region planning and demand characteristics. Step 42: Extract the list of retailer product specifications within the regional planning scope of this cluster, and for each product specification that the retailer can sell. Based on its characteristics, the degree to which each requirement is satisfied is quantified, resulting in... ,in, A higher value indicates a better fulfillment of requirement k; Step 43: Calculate the overall demand score for each product to evaluate the product. The overall degree of alignment with consumer needs; Step 4: Construct the product feature vector space, and then use One-Hot encoding to process discrete variables; Steps four and five: Calculate the similarity between the product specifications sold by each retailer. .

4. The multi-level supply chain business-industry collaborative distribution method according to claim 1 or 2, characterized in that: The specific steps for adjusting the product specification distribution ratio in step five are as follows: Step 51: Based on similar products for each product specification, calculate the overall competitiveness of similar products with this product specification; Then, balance your own needs with the impact of similar products; Step 52: Under ideal conditions, calculate the placement ratio based entirely on matching consumer demand. Step 53: Integrate historical data, control the adjustment intensity, and update the allocation ratio; Step 54: Set the minimum delivery ratio and perform normalization processing; Step 55: Send advertising adjustment suggestions to enterprises.

5. A system applying the multi-level supply chain business-industry collaborative distribution method according to any one of claims 1 to 4, characterized in that: include: Data acquisition module: used to capture multi-source heterogeneous comment data and perform preprocessing; Demand Feature Mining Module: Employs deep learning models to extract multidimensional consumer demand features; Demand clustering and similar product calculation module: integrates demand characteristics and product similarity calculation to achieve micro-market segmentation; Product specification allocation ratio adjustment module: Generates alternative combination suggestions based on product attributes and demand matching.

Images