Shelf display recommendation method and device

Abstract

The invention discloses a shelf display recommendation method and device, and belongs to the technical field of retail management and business intelligence. According to the method, static data, dynamic sales data, customer behavior data, external environment data and inventory data are collected and fused, a multi-dimensional display effect evaluation index is constructed, a display strategy knowledge base is formed based on a machine learning model, a candidate scheme is generated by using an optimization algorithm and digital simulation is performed, and an optimal scheme is recommended and visually output. Meanwhile, iterative optimization of the model is realized through closed-loop feedback; the method realizes the conversion from experience driving to data driving, from static state to dynamic state real-time, and from single target to multi-target global optimization, has prediction simulation capability and self-evolution capability, and can significantly improve the scientificity and effectiveness of shelf display.

Description

Technical Field

[0001] This invention relates to the fields of retail management and business intelligence technology, and in particular to a method and apparatus for recommending shelf displays. Background Technology

[0002] In the retail industry, shelf display is a crucial point of contact between products and customers, directly impacting customer purchasing decisions and overall store sales. A scientifically designed and well-organized shelf display can effectively attract customer attention, increase product exposure, promote cross-selling, and thus significantly improve sales per square foot (sales per unit area).

[0003] However, traditional shelf display solutions have many limitations. They are generally based on subjective opinions and rely on the personal experience and intuition of store managers. They lack unified scientific standards, resulting in significant differences in display effects between different stores and making it difficult to replicate successful experiences on a large scale.

[0004] Some display decisions are often based solely on historical financial data such as sales volume and gross profit margin, ignoring a wealth of behavioral data such as customer movement patterns in the store, time spent in front of shelves, product picking and putting back behavior, and eye focus hotspots, as well as external contextual information such as weather, holidays, and the surrounding competitive environment. Furthermore, the potential effects of new display schemes cannot be predicted and evaluated before implementation, resulting in high trial-and-error costs. If the scheme is not effective, it may directly lead to a loss of sales.

[0005] Existing automated product display recommendation solutions often employ simple association rules (such as product association recommendations based on shopping basket analysis) or rankings based on historical sales data (such as placing high-selling products in good positions), which still fail to solve the core issues mentioned above, such as multi-dimensional data integration, real-time dynamic adjustment, and prediction of solution effectiveness.

[0006] Therefore, a technical solution is needed that can comprehensively utilize multi-source heterogeneous data and dynamically generate the optimal display scheme. Summary of the Invention

[0007] The purpose of this invention is to provide a shelf display recommendation method and apparatus to solve the problems in the background art.

[0008] To achieve the above objectives, the present invention provides a shelf display recommendation method, comprising the following steps:

[0009] S1. Collect multi-dimensional data related to shelf display and process the data to obtain a feature dataset;

[0010] S2. Construct a multi-dimensional display effect evaluation index based on the feature dataset;

[0011] S3. Based on historical data in the feature dataset, a machine learning model is used for training to learn the mapping relationship between product attributes, display location, customer behavior, environmental factors and display effect evaluation indicators, and to build a display strategy prediction model.

[0012] S4. Set constraints for the target shelf and use optimization algorithms to generate multiple candidate display schemes; input the features of each candidate display scheme into the display strategy prediction model for processing to obtain the comprehensive effect score of each scheme;

[0013] S5. Select the candidate solution with the highest performance score as the optimal recommended solution and generate a visual display guide map;

[0014] S6. Implement the optimal recommended plan and collect actual data after the implementation of the recommended plan. Calculate the actual effect based on the display effect evaluation index in S2, and update the display strategy prediction model based on the results.

[0015] Preferably, in S1, the multi-dimensional data includes static data, dynamic sales data, customer behavior data, external environment data, and inventory data; the static data includes product attribute data, shelf physical parameters, and store layout diagram; the product attribute data includes SKU information, category, subcategory, brand, specifications, size, weight, price, shelf life, and packaging type; the shelf physical parameters include shelf number, location, number of layers, width, depth, height, orientation, and material of each layer; the store layout diagram includes entrance, exit, cash register, aisle distribution, and division of areas for each product category;

[0016] Dynamic sales data includes real-time or near real-time product sales revenue, sales volume, average order value, conversion rate, sales changes during promotional activities, and sales curves for different time periods.

[0017] Customer behavior data consists of anonymized data collected through in-store cameras (with privacy protection measures such as anonymization and de-identification), Wi-Fi probes, Bluetooth beacons, smart shopping carts, RFID tags, or sensor devices. This includes heat maps of customer movement within the store, dwell time on specific shelves or product areas, number of times products are picked up, number of times products are put back, purchase conversion rate after picking up products, eye-tracking heat maps, and average time customers spend in the store.

[0018] External environmental data includes real-time and future weather forecasts (temperature, humidity, precipitation, wind speed), holiday information, local large-scale event information, trending topics on social media (especially discussions related to the products), and product prices and promotional information of nearby competitors, obtained through third-party APIs or web crawlers.

[0019] Inventory data includes real-time inventory levels, inventory locations (store warehouses or shelves), inventory turnover rate, replenishment cycle, replenishment in transit information, near-expiry product information, and out-of-stock records.

[0020] Preferably, in step S1, the data processing procedure is as follows: standardizing data from different sources and in different formats, for example, by unifying the timestamp format and geographic coordinate system;

[0021] Handling missing and outlier values ​​in data;

[0022] By linking customer behavior data with product sales data, the number of people stopping at a certain shelf during a certain period is matched with the sales volume of products on that shelf during that period.

[0023] Align external environmental data with store data over time, for example, link weather data with daily sales data.

[0024] Preferably, in S1, the feature dataset contains multi-dimensional features describing products, shelves, customers, environment, inventory, and sales status.

[0025] Preferably, in step S2, the display effectiveness evaluation index is an objective function that reflects the retailer's overall business objectives, specifically expressed as follows:

[0026] objective = α·sales revenue + β·gross profit margin + γ·inventory turnover rate + δ·risk of slow-moving goods + ε·exposure of strategic products;

[0027] In the formula, α, β, γ, δ, and ε are adjustable weighting coefficients, which sum to 1 and are set according to the business strategy.

[0028] For example, during the inventory clearance phase, γ and δ can be increased; during the promotion of new products or strategic commodities, ε can be increased; and when pursuing short-term performance, α and β can be increased.

[0029] The risk of slow-moving inventory can be calculated by comprehensively considering indicators such as the number of days of inventory backlog and the recent month-on-month decline in sales; the exposure of strategic products can be measured by indicators such as the number of times the product is looked at by customers and its display position in popular traffic areas.

[0030] Preferably, in step S3, the machine learning model can be Gradient Boosting Decision Tree (GBDT), Random Forest algorithm, Neural Network algorithm, etc. During the training process, the features of the historical display scheme, including product attributes, display location coordinates, number of cards, customer behavior characteristics, external environment characteristics at the time, and inventory characteristics, are taken as input, and the comprehensive effect score actually generated by the scheme is taken as the output label. Multiple rounds of iterative training are carried out to optimize the model parameters and minimize the prediction error of the model.

[0031] Preferably, historical data includes successful display cases, unsuccessful display cases, and their corresponding performance data.

[0032] Preferably, in step S3, the constructed display strategy prediction model uses the display plan, current external environment data, and inventory data as inputs to estimate the overall effect score of the display plan.

[0033] Preferably, in S4, the price adjustment constraint includes inventory constraints and product display restrictions, specifically including real-time inventory quantity limits, product size matching restrictions with shelf space, and product display prohibition rules (such as dangerous goods not being on the same shelf as food, competing brands not being adjacent, and promotional activity requirements). The optimization algorithm uses a genetic algorithm or Monte Carlo tree search algorithm to generate candidate display schemes under preset constraints. Each candidate scheme specifies in detail which SKUs are displayed in which specific locations on the shelf, the number of display spaces for each SKU, and the orientation of the products.

[0034] Preferably, in step S5, the optimal recommendation scheme is visualized through a graphical interface, and the display guidance diagram marks the specific display location, product orientation, suggested number of display spaces, and display instructions for each product;

[0035] Additionally, a forecast report on the projected performance of the plan can be attached, including the projected increase in various indicators such as sales revenue and gross profit margin.

[0036] Preferably, in step S5, if there are multiple solutions with similar overall performance scores, a secondary screening can be performed based on auxiliary indicators, such as implementation difficulty and adjustment range.

[0037] Preferably, in step S6, the actual effect is compared with the predicted score obtained in step S4, and the display strategy prediction model in step S3 is evaluated based on the comparison result.

[0038] The data of the optimal recommendation scheme is then used as sample data and re-inputted into step S3. The machine learning model is then used for incremental learning or full retraining to update the model parameters, thereby achieving model optimization and adaptive iteration.

[0039] The present invention also provides a shelf display recommendation device for implementing the above recommendation method.

[0040] Preferred, including:

[0041] Data Acquisition and Fusion Module: Used to execute step S1; acquire data from various internal data sources (such as ERP system, POS system, WMS warehouse management system, store camera management system, customer behavior analysis system) and external data source interfaces (such as weather API, social media data interface, competitor information collection interface), and perform data processing, including data cleaning (duplicate removal, noise reduction, missing value filling), format conversion, feature extraction and data fusion, to generate a standardized feature dataset for subsequent processing;

[0042] Core computing engine module: Used to execute steps S2, S3, and S4, it is the core processing unit of the device, including:

[0043] Objective Function Management Unit: Used to define, store, modify, and load comprehensive display effect evaluation indicators. Allows users to adjust the weight coefficients of each indicator (sales, gross profit margin, etc.) in the objective function of the evaluation indicators according to different business goals and scenarios, and provides preset weight templates (such as inventory clearance templates and new product promotion templates).

[0044] Machine Learning Model Unit: Used to build, train, evaluate, store and load display strategy prediction models. It supports the selection and configuration of various machine learning algorithms (such as GBDT, random forest, neural network). It can receive historical feature data and corresponding effect labels provided by the data acquisition and fusion module for model training, and can receive new feedback data for model updates (incremental learning or retraining).

[0045] The optimization simulation unit is used to generate candidate display schemes and perform effect prediction simulations. Based on the target shelf range specified by the user and the current constraints (obtaining real-time inventory and other information from the data acquisition and fusion module), it uses optimization algorithms to generate a large number of feasible candidate display schemes. Each candidate scheme is converted into a feature vector and, combined with the current environmental features, is input into the prediction model in the machine learning model unit to obtain the prediction effect score of each scheme.

[0046] Recommendation and Interaction Module: Used to execute step S5; provides a graphical user interface (GUI) to enable user interaction with the system, allowing users to input target shelves, set constraints, select weight templates for comprehensive display effect evaluation indicators, or customize weights; displays the predicted effect ranking of candidate solutions to users; displays the best solution in a visual display guidance diagram (2D or 3D), supporting zooming, zooming, and rotation, and can be exported as an image or PDF for printing; simultaneously displays a predicted effect report of the solution.

[0047] Feedback Learning Module: Used to execute step S6; monitor the implementation status of the recommended plan (which can be manually marked by the user or automatically associated with the implementation time); after the plan is implemented, within a preset period (e.g., 1 day, 3 days, 1 week), automatically obtain newly generated sales data, customer behavior data, etc. from the data collection and fusion module, call the objective function management unit to calculate the actual effect score of the plan; combine "display plan - actual effect - environment and inventory characteristics at the time of implementation" into new training samples, store them in the historical plan library, and trigger the model update process of the machine learning model unit to use the new samples for model retraining, so as to realize the continuous evolution of the model;

[0048] Data storage module: Used to store the data in steps S1-S6, and to persistently store various types of data, including: raw data (unprocessed data collected from various data sources), fused feature data (output of the data collection and fusion module), historical scheme library (all historical display schemes and their actual effect data), model parameters (model parameters and configuration information obtained from training machine learning model units), system configuration information (such as user accounts, permission settings), etc.

[0049] Therefore, the shelf display recommendation method and apparatus of the present invention have the following beneficial effects:

[0050] (1) This invention comprehensively utilizes multi-dimensional and multi-source data such as commodities, sales, customer behavior, environment, and inventory, and mines the potential correlation between data through artificial intelligence models, so that display decisions are freed from excessive reliance on personal experience and are more scientific, objective and accurate; and can collect and process data in real time or near real time, and quickly respond to changes in the internal and external environment, such as inventory fluctuations, sudden weather changes, social media hotspots, promotional activities, etc., to realize on-demand and timely dynamic optimization and adjustment of display plans, and improve the timeliness of decision-making.

[0051] (2) This invention balances multiple key performance indicators such as sales revenue, gross profit margin, inventory turnover rate, risk of slow-moving goods, and promotion of strategic products by using a customized comprehensive display effect evaluation index, thereby avoiding the suboptimal global result that may be caused by optimizing a single indicator and achieving the global optimal retail goal.

[0052] (3) Before the implementation of the scheme, the present invention uses the display strategy knowledge base to predict and simulate the effect of the candidate schemes, so that retailers can know the potential effect of each scheme in advance and select the optimal scheme, which greatly reduces the trial and error cost and the opportunity loss caused by the inappropriate scheme; in addition, through the closed-loop feedback mechanism, the system can continuously learn new knowledge and experience from each display practice, continuously optimize the machine learning model, and make its predictive ability and recommendation effect continuously improve with the increase of usage time.

[0053] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0054] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0055] Figure 2 This is a diagram illustrating the data processing procedure according to an embodiment of the present invention. Detailed Implementation

[0056] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0058] Example

[0059] Taking the autumn apparel display optimization of a fast-fashion chain store as an example, the store hopes to increase sales of new autumn products, clear remaining summer inventory, and drive store traffic by optimizing the display of a certain trendy sweatshirt (a popular item on social media). Figure 1 , Figure 2 As shown, the specific steps include:

[0060] S1: Multi-source data acquisition and fusion: Collect multi-dimensional data related to shelf display and process the data to obtain a feature dataset; specifically: the data acquisition and fusion module initiates full-dimensional data acquisition, as follows:

[0061] Static data: Obtain apparel SKU information from the product database, including detailed parameters of autumn new arrivals (such as knitted cardigans and cargo pants), summer slow-moving items (such as short-sleeved T-shirts and shorts), and trendy hoodies—style (hooded / crew neck), material (cotton / polyester), size (S / M / L / XL), color (black / white / grey / Morandi color scheme), price range (99-299 yuan), and seasonal attribute (autumn / summer); Obtain shelf parameters from the store system: the main aisle has golden hanging rods on both sides (1.6m high), the fitting room area is a stacking area (1.2m×0.8m), and the entrance is an endcap promotional area (1.8m wide).

[0062] Dynamic sales data: Sales records over the past 30 days show that 5-8 pieces of autumn knitted cardigans are sold per day, while less than 2 pieces of summer short-sleeved T-shirts are sold per day; 20 pieces of the popular sweatshirts have sold out in 3 days and are being restocked.

[0063] Customer behavior data: collected through in-store cameras (anonymized) and smart fitting room system: customers spend an average of 45 seconds in the hoodie area, with a 30% try-on conversion rate; the summer short-sleeved area has a short dwell time (<10 seconds), with an 80% return rate after picking up items; clothing on the prime hanging rods is touched 3 times more often than clothing on the bottom shelves.

[0064] External environmental data: Weather forecasts indicate that temperatures will drop to 15-22℃ in the next 10 days, making it suitable for wearing knitted cardigans; social media monitoring shows that the combination of "oversized hoodie + cargo pants" has become a popular topic among students; competitors in the area are charging 20 yuan more for the same style of hoodie.

[0065] Inventory data: 500 summer short-sleeved T-shirts are still in stock (15 days left until the end-of-season clearance), there is ample stock of autumn knitted cardigans, and 200 trendy hoodies will be restocked and arrive in stores in 2 days.

[0066] The data fusion module cleans and correlates the above data, and the data fusion results in the following characteristics: the sales conversion rate of the combination of "temperature 15℃ + hooded sweatshirt + golden hanging rod" is high; the trial rate of the popular sweatshirt increases by 50% during the "student gathering time (18:00-21:00)".

[0067] S2. Quantification and Target Definition of Display Effectiveness: Construct a multi-dimensional display effectiveness evaluation index based on the feature dataset; specifically:

[0068] Store operations managers log into the system via the recommendation and interaction module. On the system interface, they set the objective function for evaluating display effectiveness based on current business goals, with weights biased towards inventory clearance and new product promotion.

[0069] Objective = 0.25 × Sales Revenue + 0.2 × Gross Profit Margin + 0.3 × Inventory Turnover Rate (Summer Style) - 0.15 × Risk of Unsold Summer Styles + 0.1 × Exposure of Influencer Hoodies;

[0070] Increase inventory turnover rate weight (0.3) to accelerate summer clearance; reduce unsold inventory risk weight (0.15) to avoid summer stock backlog; retain the exposure weight of trendy hoodies (0.1) to drive customer traffic;

[0071] S3. Display Strategy Model Construction: Based on historical data in the feature dataset, a machine learning model is trained to learn the mapping relationship between product attributes, display location, customer behavior, environmental factors, and display effect evaluation indicators, thus constructing a display strategy prediction model; specifically:

[0072] The machine learning model unit calls a random forest model trained on two years of historical clothing display data. During this process, the ratio of training set, validation set, and test set is 7:2:1. After model training, key patterns are learned:

[0073] When the temperature is between 16-20℃, the sales of autumn knitted cardigans displayed on the main aisle's prime hanging rods (1.5-1.7m) are 2.3 times that of the bottom shelves;

[0074] When trending items are displayed next to basic items (such as white T-shirts), cross-purchase rates increase by 40%.

[0075] Summer slow-moving items are offered with a "buy one get one free" promotion and displayed in the stacked area at the entrance of the fitting rooms, which speeds up the clearance process by 3 times.

[0076] During peak student hours (after 6 PM), oversized hoodies were moved to the end racks near the entrance, increasing the fitting rate by 65%.

[0077] By using the aforementioned key principles as a knowledge base for display strategies, we can predict the effectiveness of new display plans and provide core decision-making basis for subsequent steps.

[0078] S4. Candidate Solution Generation and Digital Simulation: For the target shelf, constraints are set, and multiple candidate display solutions are generated using optimization algorithms. The features of each candidate display solution are input into the display strategy prediction model for processing, yielding a comprehensive score for each solution. Specifically:

[0079] The simulation unit was optimized to set constraints for the three core areas of the store (main aisle hanging rods, fitting room stacking area, and entrance end shelf): popular hoodies need to reserve 20% of the hanging rod space, summer styles need to be displayed together, and clothing of the same size should be placed adjacent to each other.

[0080] 300 candidate solutions are generated using a genetic algorithm. The optimization simulation unit inputs each candidate solution into the display strategy knowledge base and, combined with the current environmental characteristics, predicts the comprehensive effect score (Objective value) of each solution.

[0081] Option A: Display knitted cardigans (alternating black and gray) on hanging rods along the main aisle; set up "buy one get one free" displays of summer short-sleeved shirts in the fitting room stacking area; and display trendy hoodies paired with cargo pants on endstands at the entrance. The predicted overall score for the next 3 days is 89 points, summer inventory can be reduced by 30%, and sales of the trendy hoodies are expected to reach 50 pieces.

[0082] Option B: Popular hoodies occupy prime main aisle racks, while summer styles are moved to corner shelves. Predicted score: 76 points; sales are expected to decrease by 15% due to poor display placement of cardigans.

[0083] Option C: No styling options are displayed; the popular hoodie is displayed separately. Predicted score: 82 points; cross-purchase rate is 12% lower than Option A.

[0084] All other options are inferior to option A, and will not be described in detail here.

[0085] S5. Optimal Solution Recommendation and Visualization Output: Select the candidate solution with the highest performance score as the optimal recommended solution and generate a visual display guide; specifically:

[0086] The recommendation and interaction module outputs solution A as the optimal solution, generating a 3D display guidance diagram:

[0087] Main aisle hanging rod: black knitted cardigans (SL size) are hung in columns 1-3, and gray knitted cardigans (M-XL size) are hung in columns 4-6. The hooks are spaced 8cm apart to ensure visual neatness.

[0088] Fitting room stacking area: Labeled with red tags "Buy one get one free on summer short sleeves", stacking height not exceeding 30cm, 10 pieces per layer (2 pieces each of sizes S-XL);

[0089] Entrance end shelf: On the left hang trendy hoodies (with hooded ones in the center), on the right hang cargo pants, and in the middle place a mannequin displaying the complete outfit, labeled "Scan the QR code to view the outfit video";

[0090] Simultaneous output forecast report: Sales are expected to increase by 22% on the 3rd, summer short-sleeved shirt inventory will drop to 350 pieces, and trendy hoodies will drive an 18% increase in customer traffic.

[0091] S6. Closed-Loop Feedback and Model Iteration: Implement the optimal recommended solution and collect actual data after its implementation. Calculate the actual effect based on the display effect evaluation indicators in S2, and update the display strategy prediction model based on the results. Specifically: Store staff implement solution A according to the display guidance diagram. Three days after the implementation of solution A, the feedback learning module collects actual data.

[0092] Actual sales increased by 25%, exceeding forecasts by 3 percentage points;

[0093] Summer short-sleeved shirt inventory decreased by 320 pieces (clearance rate 64%), better than expected;

[0094] The associated purchase rate of trendy hoodies and cargo pants reached 38%, compared to a predicted 30%, due to the significant effect of models pairing them together;

[0095] The conversion rate of trying on hoodies after 6 PM, when they were moved to the endcaps, reached 45%, compared to a predicted 35%.

[0096] The feedback learning module calculates the actual effect score of Scheme A and stores the data pair of "Scheme A parameters - actual effect - cooling weather + student passenger flow characteristics" into the historical scheme database as new training samples. The system triggers the incremental learning process of the machine learning model unit, using the new samples to update the model. The updated model strengthens the feature weights of "scenario-based matching display" and "time-based display adjustment", providing more accurate predictive capabilities for subsequent winter clothing display optimization.

[0097] Therefore, the present invention provides a shelf display recommendation method and apparatus that collects and integrates static data, dynamic sales data, customer behavior data, external environment data, and inventory data to construct multi-dimensional display effect evaluation indicators. Based on a machine learning model, a display strategy knowledge base is formed. Candidate solutions are generated using optimization algorithms and digitally simulated. The optimal solution is recommended and visualized. Simultaneously, closed-loop feedback enables iterative optimization of the model. This achieves a transformation from experience-driven to data-driven, from static to dynamic real-time, and from single-objective to multi-objective global optimization. It possesses predictive simulation capabilities and self-evolutionary capabilities, significantly improving the scientific rigor and effectiveness of shelf displays.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for recommending shelf displays, characterized in that, Includes the following steps: S1. Collect multi-dimensional data related to shelf display and process the data to obtain a feature dataset; S2. Construct a multi-dimensional display effect evaluation index based on the feature dataset; S3. Based on historical data in the feature dataset, a machine learning model is used for training to learn the mapping relationship between product attributes, display location, customer behavior, external environment and display effect evaluation indicators, and to build a display strategy prediction model. S4. Set constraints for the target shelf and use optimization algorithms to generate multiple candidate display schemes; input the features of each candidate display scheme into the display strategy prediction model for processing to obtain the comprehensive effect score of each scheme; S5. Select the candidate solution with the highest performance score as the optimal recommended solution and generate a visual display guide map; S6. Implement the optimal recommended plan and collect actual data after the implementation of the recommended plan. Calculate the actual effect based on the display effect evaluation index in S2, and update the display strategy prediction model based on the results.

2. The shelf display recommendation method according to claim 1, characterized in that: In S1, the multi-dimensional data includes static data, dynamic sales data, customer behavior data, external environment data, and inventory data; the static data includes product attribute data, shelf physical parameters, and store layout diagram.

3. The shelf display recommendation method according to claim 1, characterized in that: In S1, the data processing procedure is as follows: standardize data from different sources and in different formats; Handling missing and outlier values ​​in data; Link customer behavior data with product sales data; Align external environment data with store data over time.

4. The shelf display recommendation method according to claim 1, characterized in that: In S2, the evaluation index for display effectiveness is expressed as follows: objective = α·sales revenue + β·gross profit margin + γ·inventory turnover rate + δ·risk of slow-moving goods + ε·exposure of strategic products; In the formula, α, β, γ, δ, and ε are weighting coefficients, which are set according to the business strategy.

5. The shelf display recommendation method according to claim 1, characterized in that: In S3, during the training process of the machine learning model, the features of the historical display scheme, the environmental features at that time, and the inventory features are taken as inputs, and the comprehensive effect score of the corresponding scheme is taken as the output label. Iterative training is carried out to optimize the model parameters and minimize the prediction error of the model.

6. The shelf display recommendation method according to claim 1, characterized in that: In S4, the price adjustment constraint includes inventory constraints and rules on products that cannot be displayed; the optimization algorithm uses a genetic algorithm or a Monte Carlo tree search algorithm to generate candidate display schemes under preset constraints.

7. The shelf display recommendation method according to claim 1, characterized in that: In S5, the optimal recommendation scheme is visualized through a graphical interface. The display guidance diagram marks the specific display location, product orientation, suggested number of display spaces, and display instructions for each product.

8. The shelf display recommendation method according to claim 1, characterized in that: In step S6, the actual effect is compared with the predicted score obtained in step S4, and the display strategy prediction model in step S3 is evaluated based on the comparison results. The data of the optimal recommendation scheme is then used as sample data and re-inputted into step S3. The machine learning model is then used for incremental learning or full retraining to update the model parameters, thereby achieving model optimization and adaptive iteration.

9. A shelf display recommendation device, characterized in that: Used to implement the recommended method as described in any one of claims 1-8.

10. The apparatus according to claim 9, characterized in that: include: Data acquisition and fusion module: used to execute step S1; Core computing engine module: used to execute steps S2, S3, and S4; Recommendation and Interaction Module: Used to execute step S5; Feedback learning module: used to execute step S6; Data storage module: Used to store the data in steps S1-S6.

Images