LLM guides NSGA-III multi-objective product selection method and system
By using the LLM-guided NSGA-III multi-objective optimization method combined with a large language model, the problems of multi-objective conflict and data processing in existing product selection technologies are solved. This enables product role recognition and the generation of interpretable product selection reports, improving the executability and repeatability of product selection results.
Patent Information
- Application Number
- CN202611294881.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-25
- Publication Date
- 2026-11-10
AI Technical Summary
Existing product selection technologies struggle to simultaneously maximize total profit, minimize the risk of unsold inventory, maximize category coverage, and balance the product role structure. They also lack the ability to automatically clean heterogeneous fields, construct derived indicators, identify product roles, generate Pareto schemes, and sort interpretable preferences. Traditional random initialization leads to a large number of infeasible solutions, reducing search efficiency. Product selection results rely on human experience, have poor repeatability, unstable combination structures, and are difficult to implement.
By adopting the LLM-guided NSGA-III multi-objective optimization method and combining it with a large language model, the system automatically cleans data, identifies product roles, constructs a multi-objective combined model, and generates an interpretable product selection report through data access, preprocessing, product role identification, multi-objective model construction, NSGA-III optimization solution, TOPSIS ranking, and result output. This solves the above problems.
It achieves an effective balance between multiple objectives, automatically adapts to heterogeneous data structures, identifies product roles, generates interpretable product selection reports, reduces data preparation costs, and improves search efficiency and the interpretability and executability of product selection results.
Smart Images

Figure CN123070404A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of commodity management, specifically to a commodity selection method and system that combines a large language model, evolutionary optimization algorithm, and multi-objective optimization method. Background Technology
[0002] With the development of e-commerce, content e-commerce, and instant retail, product selection has gradually shifted from being driven by traditional experience to being driven by data. For products such as tea drinks, food, health drinks, and daily chemical fast-moving consumer goods, which have differences in shelf life, specifications and net content, cost constraints, and brand positioning, product selection decisions not only need to determine whether an individual product is worth selling, but also need to consider total profit, risk of unsold inventory, category coverage, and the proportion of marketing roles at the portfolio level.
[0003] Existing product selection methods mainly include manual experience screening, historical sales ranking, single-indicator gross profit ranking, simple weighted scoring, or recommendation models based on sales forecasts. While these methods can reduce manual workload to some extent, they still have significant shortcomings. First, single-indicator ranking struggles to handle conflicts between multiple objectives. High-profit products may have higher prices and slower sales; low-priced products may be suitable for attracting customers but have limited profit contribution; products with short remaining shelf lives, even with higher profits, face the risk of near-expiration and inventory depreciation. Simply merging multiple objectives into a fixed score can easily mask the true conflicts between them.
[0004] Secondly, product selection is essentially a combination optimization problem, not an isolated problem of ranking individual products. Without product role identification and control over the proportion of combined roles, the product selection results may lead to an unbalanced product structure. At the same time, existing role classification methods often introduce too many classification dimensions and judgment conditions, and there are overlaps and omissions between rules. Additional fallback rules are needed to cover all products, resulting in blurred role boundaries and difficulty for business personnel to understand intuitively.
[0005] Secondly, products have a clear shelf-life constraint. Traditional methods often use shelf-life as a simple filtering condition, failing to model the remaining shelf life in conjunction with product cost, making it difficult to accurately reflect the portfolio risk of high-cost, near-expiry products.
[0006] Furthermore, while multi-objective evolutionary algorithms can output multiple Pareto non-dominated solutions, without a subsequent comprehensive ranking mechanism, the Pareto solution set is difficult to directly transform into an executable product list. In addition, traditional genetic algorithms use random methods to generate the initial population; randomly generated 0-1 vectors often fail to simultaneously satisfy multiple constraints such as the range of SKU numbers, core category coverage, and role structure matching. A large number of infeasible solutions cause the algorithm to primarily focus on fixing constraint violations in early iterations, thereby reducing overall search efficiency.
[0007] Finally, most existing multi-objective product selection systems only output numerical indicators and product lists, lacking natural language interpretation of optimization solutions and output of implementation strategies. Business personnel need to analyze the logic of the solutions and formulate operational actions on their own, which has high costs for understanding and implementation.
[0008] Therefore, there is a need for a method and system that can automatically clean candidate product data, identify product business roles, build multi-objective combination product selection models, improve efficiency by using LLM-guided initialization, and output the final solution and generate an interpretable report through collaboration between NSGA-III and TOPSIS. Summary of the Invention
[0009] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an LLM-guided multi-objective product selection method and system for NSGA-III. The problems to be solved are: existing product selection technologies struggle to simultaneously maximize total profit, minimize the risk of unsold inventory, maximize category coverage, and balance the product role structure; they also lack the ability to automatically clean heterogeneous fields, construct derived indicators, identify product roles, generate Pareto schemes, and rank interpretable preferences; traditional random initialization leads to a large number of infeasible solutions, reducing search efficiency, resulting in product selection results that rely on human experience, have poor repeatability, unstable combination structures, and are difficult to implement.
[0010] The method includes: data access S100, preprocessing S200, product role recognition S300, multi-objective model construction S400, NSGA-III optimization solution S500, TOPSIS ranking S600, and decision output S700. The system includes corresponding functional modules: data access module 100, preprocessing module 200, product role recognition module 300, multi-objective model construction module 400, NSGA-III optimization solution module 500, TOPSIS ranking module 600, large language model interaction module 700, and result output module 800.
[0011] This invention also provides a multi-objective optimization product selection decision-making system for implementing the above-described method. The system includes a data access module 100, a preprocessing module 200, a product role recognition module 300, a multi-objective model construction module 400, an NSGA-III optimization solution module 500, a TOPSIS ranking module 600, a large language model interaction module 700, and a result output module 800. Each module is used to execute steps S100 to S700. The system also includes an interactive display layer, providing users with functions such as uploading candidate product libraries, setting weights, setting evolutionary parameters, starting optimization tasks, displaying running status, displaying Pareto frontiers, and downloading product lists.
[0012] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention models product selection as a four-objective constraint optimization model that includes maximizing total profit, minimizing the risk of unsold goods, maximizing category coverage and minimizing the deviation of product matrix structure. The model is solved by the NSGA-III algorithm, which effectively identifies and balances multiple conflicting business objectives and overcomes the shortcomings of traditional single index ranking in handling multi-objective conflicts.
[0013] (2) This invention uses a mapping mechanism that combines a set of preset standard fields with keyword fuzzy recognition to automatically adapt to the structure of heterogeneous candidate product tables and automatically complete net content extraction, price cleaning, remaining shelf life calculation and derived index construction, thereby reducing data preparation costs.
[0014] (3) This invention uses three core dimensions—bottom-line gross profit margin, relative price coefficient, and cost-effectiveness—to automatically identify candidate products as lead-generation products, profit-generating products, best-selling products, and image-building products using a sequential judgment rule. The rules are transparent and configurable, ensuring that the product selection results pursue optimal numerical values while also taking into account the rationality of the marketing mix structure and the differences in brand positioning, and directly correspond to the inherent understanding of the product matrix by business personnel.
[0015] (4) The present invention adopts a hybrid initialization strategy that combines LLM guided rule set with random generation. LLM generates greedy scoring rules based on business semantics, which are oriented towards profit, risk, coverage and role matching. This guides part of the initial population to be efficiently distributed in the feasible domain, taking into account the quality and diversity of the initial solution set, and reducing the constraint repair overhead in the early iteration.
[0016] (5) The present invention uses the Das-Dennis method to generate a four-dimensional uniform reference direction, and combines the reference direction association and non-dominated sorting mechanism to select the environment, so that the Pareto candidate scheme set has good distribution uniformity and is suitable for high-dimensional multi-objective combination optimization scenarios with more than three objectives.
[0017] (6) Based on the Pareto candidate solution set generated by NSGA-III, this invention introduces the TOPSIS method to separate the benefit-type indicators and cost-type indicators into models, calculates the closeness of each solution to the positive and negative ideal solutions and sorts them, and outputs a unique list of preferred options that can be executed while retaining the multi-objective trade-off space.
[0018] (7) This invention uses a large language model to interact and automatically reads the final solution data in the decision output stage, generating a structured natural language report that includes a comprehensive evaluation of the solution, role structure analysis, inventory risk warning and batch shelf suggestions, thereby reducing the understanding and execution costs for business personnel.
[0019] (8) This invention incorporates the remaining shelf life and standard order quantity into the modeling of the risk of slow sales, and uses the inventory pressure within the unit remaining shelf life as the risk measurement scale, which effectively suppresses the tendency of high-cost near-expiry products to be selected, and is suitable for product selection scenarios with shelf life attributes. Attached Figure Description
[0020] Figure 1 This is a flowchart of the multi-objective product selection optimization method of the present invention; Figure 2 This is a schematic diagram of the multi-objective product selection optimization system of the present invention.
[0021] Figure 3 This is a schematic diagram of the front-end interactive interface provided in an embodiment of the present invention; Figure 4 This is a flowchart of the candidate product data preprocessing and role recognition process of the present invention; Figure 5 This is a flowchart of the multi-objective product selection optimization process based on NSGA-III of the present invention; Figure 6 This is a flowchart of the TOPSIS sorting process of the present invention; Detailed Implementation
[0022] Example 1: Overall Methodology and System Structure The method flow provided in this embodiment includes: S100, Data Access: Receives a candidate product library in CSV or Excel format, as well as the minimum number of products selected by the user. , the maximum number of selections Minimum number of product categories covered Maximum number of product categories covered Minimum category depth Evolutionary parameters and target weights .
[0023] S200, Preprocessing: Perform field standardization mapping, numerical cleaning, date parsing, removal of expired goods, and calculation of derived indicators.
[0024] S300, Product Role Recognition: Based on derived indicators, candidate products are divided into traffic-driving products, profit-generating products, best-selling products, and image-building products.
[0025] S400, Multi-objective Model Construction: Using Boolean Variables To indicate the product selection status, a four-objective constrained optimization model is constructed.
[0026] S500 and NSGA-III optimization solutions: An initial population is generated using an LLM-guided rule set and a random hybrid initialization strategy, and a Pareto candidate solution set is output through evolutionary optimization.
[0027] S600 and TOPSIS ranking: Ranking Pareto candidate solutions comprehensively according to user weights to determine the optimal solution.
[0028] S700, Decision Output: Outputs the optimal solution indicators, product list and alternative solutions, and calls an external large language model to generate a product selection decision interpretation report based on the final product selection solution data.
[0029] like Figure 1 and Figure 2 As shown, the above method is executed by a multi-objective product selection optimization system, which includes a data access module 100, a preprocessing module 200, a product role recognition module 300, a multi-objective model construction module 400, an NSGA-III optimization solution module 500, a TOPSIS ranking module 600, a large language model interaction module 700, and a result output module 800. The system is equipped with a front-end interactive interface (…). Figure 3 It supports uploading candidate product libraries, setting weights, configuring evolution parameters, starting tasks, displaying Pareto frontiers, and downloading lists. Specifically, the data access module 100 executes step S100, the preprocessing module 200 executes step S200, the product role recognition module 300 executes step S300, the multi-objective model construction module 400 executes step S400, the NSGA-III optimization solution module 500 executes step S500, the TOPSIS ranking module 600 executes step S600, and the result output module 800, in conjunction with the large language model interaction module 700, executes step S700.
[0030] Example 2: Data Preprocessing The preprocessing module 200 first performs field standardization mapping. The system pre-defines the mapping relationship between standard fields and candidate fields. For example, the cost field can be matched with "product cost", "cost price", "cost", etc.; the current selling price can be matched with "selling price", "price", etc. When there is no exact match, fuzzy identification is performed through keywords. For example, if the column name contains "bottom line", it is identified as the bottom line selling price; if it contains "production" and "date", it is identified as the production date.
[0031] Numerical cleaning includes: extracting values from net content text (e.g., extracting 150 from "150 mL (25 mL × 6)"); removing currency symbols and units from price fields and converting them to numeric types; and calculating the number of days produced based on the production date and shelf life. Then calculate the remaining shelf life: ; like This indicates that the product has expired and has been removed by the system.
[0032] The system further calculates three derived metrics: Bottom-line gross profit margin: ; Value for money: ; Relative price coefficient: ; in Product category The current median selling price.
[0033] Example 3: Product Role Recognition like Figure 4 As shown, the product role recognition module 300 classifies roles based on bottom-line gross profit margin, relative price coefficient, and cost-effectiveness, using a sequential judgment rule. Default threshold: , , , , This represents the median price-performance ratio across all product categories. The determination logic is as follows: First, when and At that time, it is identified as a flagship product, serving as a brand positioning and premium display function.
[0034] Second, when and At that time, it was identified as a profit item and used as a core source of profit.
[0035] Third, when and and At that time, it was identified as a traffic-driving product, which attracted new customers by leveraging its price advantage.
[0036] Fourth, when and At that time, it was identified as a hot-selling item and the main focus was on high-volume sales.
[0037] Fifth, the safety net rule: if If it is, it is identified as a profit payment; otherwise, it is identified as a lead generation payment.
[0038] The default target allocation for each role is set as follows: 30% for lead generation products, 40% for profit-generating products, 20% for best-selling products, and 10% for image-building products. .
[0039] Example 4: Construction of a Multi-Objective Model The multi-objective model building module 400 models the product selection problem as a Boolean four-objective constrained optimization problem. Let the number of candidate products be... Decision variables , Indicates the first One product was selected. The four optimization objectives are as follows: (1) Total profit target (maximization): ; (2) The target for slow-moving inventory risk (minimization) is measured by the inventory pressure per unit of remaining shelf life: ; in For standard order quantity, To prevent extremely small positive numbers with a denominator of zero.
[0040] (3) Category coverage target (maximum), balancing coverage breadth and total number of products selected: ; in For category collection, For category The following is a collection of candidate products. To provide auxiliary variables for category coverage, This is the breadth-weighting coefficient.
[0041] (4) Product matrix structure deviation objective (minimization): ; in For character collection, For the role The following is a collection of candidate products. For the role The expected number of selections.
[0042] Since NSGA-III uses a minimization framework, the above objectives are uniformly transformed into minimizing the objective vector: ; The constraints include: Product selection quantity range constraints: ; Category depth constraints: ; Product category coverage constraints: ; Auxiliary variable logical consistency constraint: ; Example 5: Optimization Solution of NSGA-III like Figure 5 As shown, the NSGA-III optimization solver module 500 performs the following steps: S501. Input the standardized product matrix, objective function, constraint functions, and algorithm parameters. Recommended parameter values are shown in Table 1. Table 1: Recommended Values for Key Algorithms and Constraint Parameters
[0043] S502. Generating Reference Directions. The Das-Dennis method is used to generate uniformly distributed four-dimensional reference directions on the standard simplex. Let the number of layers be... The number of reference points is , .Pick 84 reference directions are generated at the same time. Each vector satisfies and Population size It is usually set to an integer multiple of the number of reference directions.
[0044] S503. Hybrid Initialization. An initial population is generated using a hybrid strategy of LLM-guided rule set and randomization, each accounting for approximately 50%. The LLM receives prompts containing problem semantics, variable types, objective function, constraints, and design requirements, and outputs a heuristic rule template. The system parses this template into four types of executable rules: profit maximization rules (selected in descending order of average category profit and single-item profit), risk minimization rules (selected in ascending order of slow-moving risk), coverage enhancement rules (prioritizing uncovered categories), and role matching rules (selected in descending order of role distribution deviation). The randomization component generates 0-1 vectors by randomly selecting categories and SKUs to ensure diversity. All individuals, after constraint correction, satisfy the hard constraints of depth and coverage.
[0045] S504, Evaluation and Ranking. Calculate the target vector for each individual. The Pareto hierarchy is determined by non-dominated sorting based on the amount of constraint violation and the amount of constraint violation.
[0046] S505, Reference Direction Correlation and Environment Selection. First, construct the ideal point. ,in For the first The individual in the first Function values on each target Construct a hyperplane by selecting the extreme points of each target axis and calculate the intercept. Normalization: After normalization, individuals are associated with the reference direction with the smallest vertical distance. When selecting an environment, individuals with fewer associated counts are prioritized to ensure uniform distribution.
[0047] S506, Genetic Operators and Constraint Repair. Two-point crossover (probability-based) is employed. ) and bit flip mutation (probability) Generate offspring. Constraint repair is divided into four layers: the first layer repairs the range of selected product quantities, below... Then randomly add 1, if higher Then randomly set to 0; the second layer of repair category depth, for those not reaching... Randomly replenish SKUs within each product category; the third layer of repair covers a wider range of product categories than... Then randomly supplement the product categories, higher than Then, the category is randomly removed; the fourth layer repairs the consistency of auxiliary variables.
[0048] S507. Termination Judgment. After reaching the preset number of evolution generations, output the set of non-dominated candidate solutions that satisfy the constraints.
[0049] Example 6: TOPSIS Sorting like Figure 6 As shown, the TOPSIS sorting module 600 performs the following steps: S601, Construct the evaluation matrix of the scheme. Let there be a total of... The Pareto candidate scheme, the... The evaluation vector for each solution is: .
[0050] S602, Vector Normalization: ; S603, Weighted Processing: ; S604. Determine the positive and negative ideal solutions. and As a benefit-oriented indicator, and This is a cost-based indicator. (Positive ideal solution) Composed of the maximum value of the benefit-type indicator and the minimum value of the cost-type indicator; negative ideal solution It consists of the minimum value of the benefit-type indicator and the maximum value of the cost-type indicator.
[0051] S605, Calculate distance and proximity: ; ; ; according to Sort by size from largest to smallest, and the one with the highest similarity is the optimal solution.
[0052] Example 7: Large Language Model Interaction and Result Output The large language model interaction module 700 plays a dual role. During the optimization and solution phase, the module acts as an initialization strategy generator, assembling problem semantics, constraint features, and business rules into standardized prompt words and sending them to the external large language model. It then returns a heuristic rule set to guide the initial population distribution. In the decision output phase, the module only reads the final solution data after TOPSIS ranking (including four target values, role proportions, category coverage, TOPSIS proximity, and product details), assembles it into structured prompt words, and calls the large model to generate a structured natural language report containing a comprehensive solution evaluation, role structure analysis, inventory risk warning, phased shelf placement suggestions, and review observation indicators. Throughout the process, it does not write back to the source or modify intermediate data.
[0053] The output module 800 outputs the comprehensive optimal solution number, the number of selected products, the number of covered product categories, the four target values, the TOPSIS closeness, the product list Excel, and the Pareto frontier plot, and merges the decision interpretation report generated by the large language model into a hybrid analysis report.
[0054] Example 8: Application Case Taking a health and wellness e-commerce platform as an example, the candidate product library contains 1200 SKUs across six categories: health-promoting beverages, healthy snacks, nourishing dried goods, functional tea bags, ready-to-eat stews, and flavored teas. After preprocessing, 1156 valid candidate SKUs are identified. User configuration parameters are shown in Table 2.
[0055] Table 2 Product Selection Task Parameter Configuration Table
[0056] After 200 generations of evolution, the system output 49 Pareto non-dominated candidate solutions. The TOPSIS ranking of the top 8 solutions is shown in Table 3.
[0057] Table 3. Ranking of the top 8 Pareto candidate solutions
[0058] Option S1 is the optimal solution, and its list of selected products is shown in Table 4.
[0059] Table 4. Selected Products List for Scheme S1 (Excerpt)
[0060] The large language model interaction module generates a structured product selection decision interpretation report based on the S1 solution data. The core recommendations include: adopting S1 as the priority combination under the current weight; testing the products in three batches according to total profit, shelf life, and category differences; using traffic-driving products to attract new customers, promoting profit-generating products, and using image-building products for gift box scenarios; adopting small-batch warehousing and FIFO mechanisms for products with high risk of slow sales; and setting conversion targets for each category.
[0061] Example 9: Electronic Devices and Storage Media The present invention also provides an electronic device, including a processor, a memory, a network interface and a display interface, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1 to 5.
[0062] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method as described in any one of claims 1 to 5.
[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A multi-objective product selection method guided by LLM for NSGA-III, characterized in that, Includes the following steps: S100, Data Access Steps: Receive data containing... A candidate product library for each candidate product, and the minimum number of products configured by the user. , the maximum number of selections Minimum number of product categories covered Maximum number of product categories covered Category minimum depth Product Role Target Percentage Evolutionary parameters and target weights Each candidate product has a cost. Current price Bottom-line price Net content Standard order quantity Shelf life Production date and product category The candidate products can be classified into product role types. , For all categories, A collection of product roles; S200, Preprocessing Steps: Perform field standardization mapping, numerical cleaning, date parsing, and missing value handling on the candidate product database, based on the number of days since production. Calculate the remaining shelf life ,like Then the candidate product is removed, and the product derivation index is calculated; S300, Product Role Recognition Step: Based on the product derivation indicators, divide the candidate products into traffic-driving products, profit-generating products, best-selling products, and image-building products; S400, Multi-objective model construction steps: Define Boolean decision variables , Indicates the first Several candidate products were selected, and it was set that... Define the decision vector; define category coverage auxiliary variables. , Indicates product category At least one item is selected; let's assume... For category The following is a collection of candidate products. For the role The following is a collection of candidate products, and the role The target percentage is Expected number of selections , To prevent extremely small positive numbers with a denominator of zero, Let the breadth weights be the coefficients; construct the following objective function and transform it into minimizing the objective vector: Total profit target: ; Risk of slow-moving inventory: ; Category coverage target: ; Product matrix structure deviation target: ; Minimize the objective vector: ; The constraints include: Product selection quantity range constraints: ; Category depth constraints: ; Product category coverage constraints: ; Auxiliary variable logical consistency constraint: ; ; S500 and NSGA-III optimization solution steps: NSGA-III is used to perform evolutionary optimization on the four-objective selection model. Non-dominated sorting and environment selection are performed based on the reference direction. An initial population is generated by using an LLM guided rule set and a random mixed initialization strategy. Offspring are generated through genetic operators and constraint repair is performed to obtain multiple Pareto candidate selection schemes that meet the constraints. S600 and TOPSIS sorting steps: Based on the user-configured target weights. The multiple Pareto candidate solutions are normalized, weighted, and the distance between positive and negative ideal solutions is calculated and the proximity is sorted to determine the comprehensive optimal solution. S700, Decision Output Steps: Output the target value, TOPSIS proximity, number of selected products, number of categories covered, alternative solutions and product list of the comprehensive optimal product selection solution, and call an external large language model to generate a product selection decision interpretation report based on the final product selection solution data.
2. The method according to claim 1, characterized in that, In the preprocessing step, the mapping relationship between standard fields and candidate field sets is pre-set during field standardization mapping. The standard fields include at least product cost, floor price, current price, shelf life, production date, product type, product number, product name, product series, packaging form, and net content. When the column names in the candidate product library do not match completely, fuzzy identification is performed through keyword matching. Calculate the bottom-line gross profit margin Cost-effectiveness and relative price coefficient : ; ; ; In the formula, For the first Product Category The current median selling price; The product role identification step is based on the bottom-line gross profit margin. Relative price coefficient and cost performance To define the dimensions, a sequential judgment rule is used to classify roles: Image style: and ; Profit: and ; Traffic-driving products: and and ; Hot Selling Item: and ; in, For high gross profit margin threshold, For low gross profit margin threshold, For high relative price coefficient threshold, For low relative price coefficient threshold, This represents the median cost-performance ratio across all product categories; for products that do not meet the above criteria, if... If it is, it is identified as a profit payment; otherwise, it is identified as a lead generation payment.
3. The method according to claim 1, characterized in that, In the multi-objective model construction step, Boolean decision variables are defined. Category coverage auxiliary variable The following targets are set for total profit, slow-moving inventory risk, product category coverage, and product matrix structure deviation: ; ; ; ; And transform it into minimizing the target vector. ; The constraints include: ; ; ; ; 。 4. The method according to claim 1, characterized in that, The NSGA-III optimization solution steps include: The Das-Dennis method is used to generate uniformly distributed reference directions based on the target quantity; An initial population is generated using a hybrid initialization strategy of LLM-guided rule set and randomization. LLM generates a heuristic rule set based on the business semantics and constraint characteristics of the multi-objective product selection problem, which guides some individuals to be distributed within the feasible region, while the remaining individuals are generated randomly. The objective function and constraint function are evaluated for individuals in the population. The Pareto hierarchy is determined by non-dominated ordination. Individuals are associated and selected based on reference direction. Genetic operators are used to generate offspring, and constraint repair is performed on the offspring to ensure that the offspring meet the constraints of the range of selected products, the depth of product categories, and the range of covered product categories. After reaching the preset number of evolutionary generations, output a set of non-dominated candidate solutions that satisfy the constraints.
5. The method according to claim 1, characterized in that, The TOPSIS sorting steps include: Assume there is a total There are Pareto candidate solutions, for each solution... Construct evaluation vector ; Perform vector normalization on each column of indicators: ; Weighted processing: ; Total profit target and category coverage target As a profitability indicator, the target of unsold goods risk Deviation target of product matrix structure As a cost-based indicator, determining the positive ideal solution and negative ideal solution Calculate the closeness: ; ; ; according to Sort the products from largest to smallest to determine the optimal comprehensive selection scheme; The method calls an external large language model: in the initial population generation stage of the NSGA-III optimization solution phase, it acts as an initialization strategy generator to generate a heuristic rule set based on the business semantics and constraint characteristics of the multi-objective product selection problem; in the decision output stage, it only reads the final solution data after TOPSIS sorting to generate a structured product selection decision interpretation report, without writing back or modifying any intermediate product selection calculation data throughout the process.
6. A multi-objective product selection system guided by LLM for NSGA-III, characterized in that, include: The data access module (100) is used to receive the candidate product library and the range of product selection quantity, category coverage, evolution parameters and target weights configured by the user; The preprocessing module (200) is used to perform field standardization mapping, numerical cleaning, date parsing and missing value handling on the candidate product library, and to calculate product derived indicators; The product role recognition module (300) is used to classify candidate products into product roles including traffic-driving products, profit-generating products, best-selling products, and image-building products based on the product derivation indicators. The multi-objective model construction module (400) is used to construct a multi-objective product selection model with the selection of candidate products as Boolean decision variables, including total profit target, unsold risk target, category coverage target and product matrix structure deviation target, and to construct product selection quantity range constraints, category depth constraints, category coverage range constraints and auxiliary variable logical consistency constraints. The NSGA-III optimization solution module (500) is used to perform non-dominated sorting genetic optimization on the multi-objective selection model based on the reference direction. It generates an initial population by using an LLM guided rule set and a random mixed initialization strategy, generates offspring through genetic operators and performs constraint repair to obtain multiple Pareto candidate selection schemes that meet the constraints. The TOPSIS sorting module (600) is used to normalize, weight, calculate the distance between positive and negative ideal solutions and sort the proximity of the multiple Pareto candidate selection schemes according to the target weights configured by the user, and determine the comprehensive optimal selection scheme. The large language model interaction module (700) is used to generate a heuristic rule set as an initialization strategy generator during the optimization solution stage, and to receive the product selection scheme data sorted by TOPSIS and generate a product selection decision interpretation report in natural language form during the decision output stage. The results output module (800) is used to output the target value, TOPSIS closeness, number of selected products, number of categories covered, alternative solutions, product list, and product selection decision interpretation report of the comprehensive optimal product selection scheme.
7. The system according to claim 6, characterized in that, When the preprocessing module (200) performs field standardization mapping, it pre-sets the mapping relationship between standard fields and candidate field sets. When the column names in the candidate product library do not match completely, it performs fuzzy identification through keyword matching; and calculates the remaining shelf life based on the production date and shelf life. ,like Then exclude it; and calculate the baseline gross profit margin. Cost-effectiveness and relative price coefficient ,in Product category The current median selling price; The product role recognition module (300) uses the bottom-line gross profit margin Relative price coefficient and cost performance To define the dimensions, a sequential judgment rule is adopted: the image style meets the requirements. and Profits are satisfied and Traffic generation funds meet the requirements. and and Bestsellers satisfy and If the above rules are not met, If it is, it is identified as a profit payment; otherwise, it is identified as a lead generation payment. The multi-objective model building module (400) defines Boolean decision variables. Category coverage auxiliary variable The following targets are set for total profit, slow-moving inventory risk, product category coverage, and product matrix structure deviation: ; ; ; ; And transform it into minimizing the target vector. ; The constraints include: ; ; ; ; 。 8. The system according to claim 6, characterized in that, The NSGA-III optimization solution module (500) is configured to: generate reference directions using the Das-Dennis method based on the target quantity; generate an initial population using an LLM-guided rule set and a random hybrid initialization strategy; evaluate the population using objective and constraint functions, determine the Pareto level through non-dominated sorting, and perform association and environment selection based on the reference directions; generate offspring using genetic operators and perform constraint repair; and output a set of non-dominated candidate solutions that satisfy the constraints. The TOPSIS ranking module (600) constructs an evaluation vector for each Pareto candidate solution and performs vector normalization and weighting processing. and As a benefit-oriented indicator, and As a cost-based indicator, it determines positive and negative ideal solutions, and calculates Euclidean distance and proximity. And sort; The large language model interaction module (700) acts as an initialization strategy generator in the optimization and solution stage, and only reads the TOPSIS final solution data to generate a structured product selection decision interpretation report in the decision output stage.
9. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein the processor executes the computer program to implement the method as claimed in any one of claims 1 to 5.
10. A computer-readable storage medium, characterized in that, It stores a computer program thereon, which, when executed by a processor, implements the method as described in any one of claims 1 to 5.
