Intelligent recommendation precision marketing method and system based on large model

Abstract

The invention discloses an intelligent recommendation precision marketing method and system based on a large model, and belongs to the technical field of artificial intelligence and retail, and the method comprises a data perception layer which is used for collecting user related information and comprises a face recognition terminal, an Internet of Things sensor, a CRM system and transaction POS data; the intelligent decision-making layer is used for performing fusion modeling and recommendation strategy generation on the sensing data and comprises a dynamic image engine, a demand prediction large model and a recommendation strategy generator; and the execution feedback layer is used for sending the recommendation result to the user in a multi-channel manner and recycling feedback data, and comprises a shop assistant PAD push module, an electronic price tag updating module and a self-service terminal guide module. The repurchase rate and customer unit price of a retail scene can be remarkably improved, and the technical problem that a traditional retail robot cannot adapt to complex requirements of customers is solved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and retail technology, specifically to an intelligent recommendation and precision marketing method and system based on a large model. Background Technology

[0002] In traditional offline retail scenarios, due to significant differences in commission rates across different product categories, employees often prioritize recommending non-essential items with higher commission rates rather than essential products that customers actually need. This distorted sales behavior directly leads to multiple negative consequences: Firstly, customers frequently encounter product recommendations that don't match their needs, prolonging their decision-making time and creating a feeling of being pressured into buying. Secondly, while stores may see a short-term increase in average transaction value, customer satisfaction and loyalty steadily decline, ultimately resulting in lower repurchase rates and damaged reputation. Summary of the Invention

[0003] The technical objective of this invention is to address the above-mentioned shortcomings by providing an intelligent recommendation and precision marketing method and system based on a large model, which can significantly improve repurchase rate and average order value in retail scenarios and solve the technical pain point that traditional retail robots cannot adapt to the complex needs of customers.

[0004] The technical solution adopted by this invention to solve its technical problem is: A method for intelligent recommendation and precision marketing based on a large model, the implementation of which includes: The data perception layer is used to collect user-related information, including facial recognition terminals, IoT sensors, CRM systems, and transaction POS data. The intelligent decision-making layer is used to fuse and model the perceived data and generate recommendation strategies, and includes a dynamic image engine, a large demand prediction model, and a recommendation strategy generator. The execution feedback layer is used to deliver recommendation results to users through multiple channels and collect feedback data. It includes a staff PAD push module, an electronic price tag update module, and a self-service terminal guidance module.

[0005] The three-layer structure is connected by a unified data channel, realizing a complete closed-loop path for data from sensory input, intelligent modeling, recommendation output to behavioral feedback.

[0006] This method deploys facial recognition and identity verification algorithms (based on models such as ArcFace and MTCNN) to accurately identify customers upon entry into the store and automatically associates them with historical purchase records and behavioral profile data as the first input for the recommendation strategy. Simultaneously, a deep learning-driven behavior prediction model is introduced to analyze multi-dimensional data such as purchase sequences, movement trajectories, and allergy preferences to construct a real-time dynamic understanding of customer needs, which is then input into the recommendation engine for inference and matching.

[0007] To enhance interaction efficiency and customer experience, service robots possess multimodal perception and interaction capabilities. Customers can interact with the robot via voice or input commands by pointing to an image interface. The robots support continuous 24 / 7 service and are capable of batch processing complex tasks such as user queries, route navigation, and promotional product descriptions, significantly reducing the cost of human intervention.

[0008] By continuously training the recommendation algorithm incrementally through a federated learning mechanism, the personalized recommendation strategy is updated in real time. Practical data shows that this method-based intelligent recommendation can increase customer dwell time and retention rate by over 40%, while reducing unit customer acquisition cost by 25%-30%. This method completely eliminates the bias of human sales tactics, objectively recommending the most suitable products, unaffected by commission levels. It also provides 24 / 7 uninterrupted service, handling massive amounts of customer inquiries and demand analysis, significantly reducing labor costs. Satisfied customers not only increase repurchase rates but also become natural promoters of the brand.

[0009] Furthermore, this method uses a dynamic demand prediction engine based on graph neural networks and time series modeling algorithms to integrate user behavior, environmental context, and historical transaction data to construct a real-time updated demand probability field, which drives the generation of personalized recommendation strategies.

[0010] Furthermore, this method supports deep fusion processing of more than 11 types of multimodal data sources, including biometric information, spatial behavior trajectories, historical transaction records, member tags, voice signals, and facial expressions. It constructs dynamic user profiles and scene-aware maps through unified encoding and context modeling. The method also extracts features and models the data through a multimodal deep fusion mechanism to build high-precision dynamic user profiles and scene-aware models.

[0011] Furthermore, regarding the data perception layer, The face recognition terminal includes: Multispectral camera: used to fuse visible light, infrared light and depth information to achieve accurate face recognition in complex environments and enhance occlusion robustness; Voiceprint acquisition array: It adopts an anti-noise pickup design, which can still stably acquire voiceprint features in noisy environments, and support customer identity verification. The IoT sensor includes: Shelf pressure sensor: with a measurement accuracy of 0.1g, it is used to sense the picking and placement of goods in real time and help judge customer dwell behavior; UWB positioning base station: Based on the TDOA (Time Difference of Arrival) algorithm, it achieves high-precision customer positioning and is used for path tracking and movement modeling; The CRM system, Integrate with the enterprise membership management system to synchronize customer identities and retrieve historical behavior data; The transaction POS data, Provides real-time data streams including product inventory and sales records to support recommendation and replenishment logic calls.

[0012] Furthermore, the demand forecasting model is based on graph neural networks and time series forecasting methods to construct a three-dimensional demand probability matrix of users, products, and scenarios; The recommendation strategy generator generates personalized recommendation scripts that match the user profile by calling the Large Language Model (LLM), with a response latency of less than 0.5 seconds; The intelligent decision-making layer includes the following implementation process: (1) Raw data input: The structured information from the data perception layer is used as the input basis, including user profiles, location trajectories, voice commands, product status, etc.; (2) Feature extraction: Perform multi-dimensional feature extraction on the original input, including numerical normalization, time-series slicing, context label completion, etc., to generate a unified vector representation; (3) Dynamic fusion: Perform fusion operation on multi-source heterogeneous data, use Transformer structure or self-attention mechanism to construct context dependency matrix, and improve the model’s ability to perceive the intrinsic relationship between different input dimensions; (4) Real-time streaming data enters the "demand forecasting model", including time series analysis network, graph neural network (GNN) or ST-GNN module, to predict users' current and near-term potential demand; (5) Static image data (such as shelf status, customer facial status, etc.) are entered into the “Long-term Evolution of Profile” module to capture customer behavior habits and trend information; (6) Strategy optimization: Based on the output results of each path, use reinforcement learning policy network or multi-objective optimization algorithm to output the optimal action suggestion (such as recommending products, navigation path, and wording generation). (7) Execution instruction generation: Form structured execution instructions and push them to various terminals (robots, shelf screens, IoT devices, etc.) in the interactive execution layer to reach end users.

[0013] Furthermore, regarding the execution feedback layer, The store clerk PAD push module synchronizes user profiles and recommended products to the service personnel's terminals, assisting them in manual guidance and intervention operations. The electronic price tag update module dynamically controls the content displayed on the shelf price tags based on inventory information and recommendation priority, enabling promotional synchronization and inventory reminders. The self-service terminal guidance module allows users to view the recommendation logic, obtain the navigation path, and complete the QR code confirmation on the terminal interface, thus realizing a closed loop of self-service. The feedback data includes product response behavior, customer emotional characteristics, and explicit rating results, which are then anonymized before being used in the training of the federated learning model.

[0014] Furthermore, the specific process for achieving intelligent recommendation and precise marketing for customers entering the store includes: Step 1: Identity Recognition and Demand Prediction: When a customer enters the store, multispectral cameras deployed at the entrance immediately start capturing facial images. These cameras operate simultaneously in the visible, infrared, and depth bands, possessing strong light suppression and weak light enhancement capabilities to ensure high-quality images are acquired even under complex lighting conditions such as backlighting and dim lighting. Simultaneously, the service robot (based on the ROS system) is activated. The robot collects the customer's facial information through a multi-camera array (RGB + depth) and calls a locally deployed facial recognition API (based on ArcFace or FaceNet algorithms) for authentication. Upon successful recognition, the system connects to the CRM database (MySQL / Redis cache) via Redis caching to match member information. If the customer is identified as a registered existing customer, a customized voice greeting is automatically generated and broadcast in real-time using a speech synthesis model (Tacotron2 + WaveRNN), enhancing user engagement and friendliness. The backend recommendation engine (TensorFlow Serving) analyzes the data across three dimensions in real time. The time series analysis module calls the Prophet model to perform time series modeling on customers' past orders and predict users' periodic consumption trends; Association rule mining (Apriori algorithm) combines customer information to extract associated products from the product knowledge graph constructed from the Neo4j graph database; The real-time inventory system queries the product status in the ERP system in real time via REST API to ensure sufficient recommended products and effective promotional activities; If this is a first-time customer, the system will guide them to register as a member via WeChat Mini Program. The WeChat Mini Program API will generate a dynamic QR code, and the form information will be automatically entered into the member information using OCR recognition (based on PaddleOCR). The registration form includes: Date picker: Used by users to enter their date of birth, which is then used to optimize subsequent recommendation strategies; Preference selector: Allows users to select personalized preferences, including product categories, price ranges, service types, allergens, etc., and the system builds recommendation filtering rules based on these preferences; The facial database encryption system (using AES-256 encryption) encrypts and encodes the facial features of family members to construct a parent-child relationship map, which is used for product recommendation and service recognition in subsequent scenarios. Step 2: Confirm and proceed to the target shelf: Recommendations are displayed on an Android interactive screen; customers confirm or modify the recommendation list via the touchscreen, and the robot activates the SLAM navigation system (LiDAR + IMU + UWB fusion positioning). Path planning uses an improved A* algorithm, and dynamic obstacle avoidance relies on the YOLOv5 real-time detection model. Upon reaching the target shelf, AR tag recognition (ARKit) assists in precise positioning, and the robot confirms the product location through dual-mode positioning (QR code + RFID). If a trial sample is required, the nearest smart sampler is activated by calling the IoT device management system (MQTT protocol) to prepare the sample. Meanwhile, customers' movement trajectories within the store are monitored in real time by an ultra-wideband positioning system, which, together with multi-view smart cameras deployed on the ceiling, forms a three-dimensional positioning network with centimeter-level precision. When a customer lingers in front of a shelf for more than a set time, and the computer vision system detects that the angle of their gaze forms a set angle with the angle of the merchandise display, the system automatically marks it as a high-intent behavior. This behavioral data is combined with the weight change sensors from the smart shelves and the results of merchandise picking and placing actions to construct a complete customer movement map. All sensor data is synchronized at the millisecond level through a precise time protocol to ensure the spatiotemporal consistency of behavioral analysis. Step 3: Scene Adaptive Processing During navigation, the system possesses the ability to automatically adapt to specific scenarios, ensuring customers receive the best service experience under different needs. It can automatically switch scene modes based on environmental perception, customer characteristics, and behavioral signals, achieving a more humanized and differentiated intelligent response, including: Environmental change response: When the system detects a significant change in the on-site environment (such as dim lighting, excessive noise, etc.), it will automatically switch to a more suitable interaction mode; for example: reducing the frequency of voice broadcasts and increasing the proportion of on-screen text prompts in noisy environments; and automatically adjusting the display brightness in strong / weak light environments to ensure visibility. User behavior state adaptation: Based on customer behavior characteristics, including dwell time, changes in walking speed, browsing frequency, etc., intelligently identify whether the user is in a state of hesitation, fatigue or high attention, and recommend rest areas, popular products, auxiliary instructions and other content to improve thoughtfulness and smoothness of experience; Multi-user identification and collaborative recommendation: When multiple customers are detected interacting at the same time (such as multiple people viewing the terminal or browsing the same area at the same time), the multi-user profiles are dynamically coordinated to provide combined recommendation solutions or adjust the interaction content synchronously to avoid information conflicts and optimize service paths. Space guidance and prompts service: Based on the store layout and the customer's current location, intelligently recommend the nearest service facilities or special areas, such as rest areas, self-checkout counters, trial areas, and information desks, to enhance the customer's sense of exploration and convenience; Step 4: Feedback Data Collection (Before Departure): Before customers leave the store, the system automatically performs a feedback data collection task, generating high-quality sample data required for closed-loop validation and model training; including: First, the actual products purchased by customers are automatically compared with the recommended list generated by the intelligent decision-making layer to calculate the recommendation hit rate, which is used to measure the accuracy of the recommendation strategy. Secondly, the facial recognition module deployed in the exit area captures the facial expression features of customers, identifies their subjective emotional state when leaving the store (such as satisfaction, confusion, fatigue, etc.), and combines it with voice tone analysis to judge potential satisfaction and interaction fluency. In addition, users are encouraged to provide proactive feedback through methods such as scanning QR codes, rating self-service terminal pages, and voice inquiries, supplementing their subjective feelings and suggestions, and building an explicit feedback data layer. All the above feedback data undergoes multiple desensitization and differential privacy algorithms before storage to ensure complete anonymization of individual identity information, retaining only the structured feature vectors used for model optimization; the feedback data will be synchronously fed into the model training module, supporting the policy engine to perform incremental learning, parameter updates, and personalized recommendation path iteration; Step 5: System Self-Evolution (Ongoing): The system has the ability to continuously learn and self-optimize, and through a federated learning architecture, it enables the online evolution of the store-level recommendation model; Each store deploys a lightweight model training engine locally to perform edge computing on customer behavior data (such as browsing paths, facial expression changes, and purchasing preferences), and only uploads encrypted model parameters to ensure data privacy and security. All uploaded parameters are aggregated to a central server, and the system performs incremental updates every hour and full model retraining every week. For new customers, a cold start strategy combining knowledge graph reasoning and best-selling product recommendations is adopted to ensure that reasonable suggestions can still be provided even without historical data. Integrate a log analysis system (such as ELK Stack) to achieve full data recording throughout the service process. All data will be uploaded to the cloud in real time, and the federated learning system will regularly update the recommendation models of each store. When a customer leaves the store, the intelligent checkout system (RFID + computer vision) automatically verifies the goods, automatically labels the purchase, updates the user profile tags, and the recommendation engine pre-generates the next service strategy.

[0015] This invention also claims a large-model-based intelligent recommendation precision marketing system, comprising: Data perception layer: This includes facial recognition terminals, IoT sensors, CRM system interfaces, and transaction POS data. It is mainly responsible for collecting raw data related to users and the environment.

[0016] Intelligent Decision Layer: Based on a dynamic image engine, a large demand prediction model, and a recommendation strategy generator, it realizes intelligent modeling and inference computation from perceived data to recommendation strategies.

[0017] Execution Feedback Layer: This layer includes multiple execution exits such as staff PAD push notifications, electronic price tag linkage, and self-service terminal guidance, responsible for quickly implementing strategy results. Simultaneously, this layer is also responsible for collecting actual user behavior data and feeding it back to the perception layer, enabling the system to continuously learn and optimize itself. The system achieves intelligent recommendation and precise marketing through the methods described above.

[0018] The present invention also claims a smart recommendation precision marketing device based on a large model, comprising: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is used to call the machine-readable program to implement the above method.

[0019] The present invention also claims a computer-readable medium storing computer instructions that, when executed by a processor, enable the implementation of the above-described method.

[0020] Compared with existing technologies, the intelligent recommendation and precision marketing method and system based on a large model of the present invention have the following advantages: This invention achieves significant technological breakthroughs and commercial value in the field of smart retail. Through an innovative multimodal biometric system, customer identification accuracy is increased to 98.5%, maintaining high robustness even in complex scenarios. The system integrates graph neural networks and real-time behavior modeling mechanisms to construct its core three-dimensional recommendation matrix algorithm (development stage × replenishment cycle × seasonal promotion), achieving 89% product recommendation accuracy, 3x improvement in replenishment reminder timeliness, and a 40% acceleration in model iteration speed through a federated learning framework.

[0021] This invention also demonstrates excellent performance in in-store interaction, navigation guidance, and path safety control. The navigation module features dynamic obstacle avoidance and path reconstruction capabilities, with emergency braking response time controlled within 200ms. Interface brightness and voice volume can be automatically adjusted based on ambient light and noise, enhancing customer comfort. These technological innovations directly translate into commercial benefits: average customer service time is reduced by 32%, with returning customers completing the recommendation interaction process in an average of 58 seconds; membership conversion rate increases by 28%, and first-time registration users increase by 45%; precise recommendations drive sales of related products, with an average increase in average order value of 35%; repurchase rate increases by 40%, and promotional product conversion rate increases by 52%.

[0022] This invention integrates precise identification, intelligent interaction, environmental adaptation, and personalized recommendations to construct an intelligent retail service ecosystem for multiple business formats and scenarios. It possesses dynamic behavior analysis and real-time strategy adjustment capabilities, enabling highly adaptable product delivery based on user status and scenario needs, reducing decision-making burden and improving shopping efficiency. Differential privacy and local encryption mechanisms ensure user data security, complying with GDPR requirements. Compared to traditional retail robots or static recommendation systems, this solution achieves a leap forward in intelligence, service experience, and business conversion rate, forming an intelligent service ecosystem with industry-leading barriers to entry. Attached Figure Description

[0023] Figure 1 This is a diagram illustrating the architecture of the intelligent recommendation and precision marketing method based on a large model provided in this embodiment of the invention. Figure 2 This is a flowchart illustrating the implementation process of the intelligent decision-making layer provided in an embodiment of the present invention; Figure 3 This is a diagram illustrating the overall structure of the intelligent recommendation and precision marketing method based on a large model provided in this embodiment of the invention. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0025] A method for intelligent recommendation and precision marketing based on a large model, the implementation of which includes: The data perception layer is used to collect user-related information, including facial recognition terminals, IoT sensors, CRM systems, and transaction POS data. The intelligent decision-making layer is used to fuse and model the perceived data and generate recommendation strategies, and includes a dynamic image engine, a large demand prediction model, and a recommendation strategy generator. The execution feedback layer is used to deliver recommendation results to users through multiple channels and collect feedback data. It includes a staff PAD push module, an electronic price tag update module, and a self-service terminal guidance module.

[0026] The demand forecasting model is based on graph neural networks and time series forecasting methods, and constructs a three-dimensional demand probability matrix of users, products, and scenarios.

[0027] The recommendation strategy generator generates personalized recommendation scripts that match the user profile by calling the Large Language Model (LLM), with a response latency of less than 0.5 seconds.

[0028] The staff PAD push module of the execution feedback layer can synchronize user profiles and recommended products to the service personnel's terminals to assist them in manual guidance and intervention. The electronic price tag update module dynamically controls the content displayed on the shelf price tags based on inventory information and recommendation priority to achieve promotion synchronization and inventory reminders. The self-service terminal guidance module allows users to view the recommendation logic, obtain navigation paths, and complete QR code confirmation on the terminal interface, realizing a self-service closed loop.

[0029] The feedback data includes product response behavior, customer emotional characteristics, and explicit rating results, which are then anonymized before being used in the federated learning model training. The system performs local preprocessing of the perceived data through edge computing nodes and issues execution instructions in a structured format after the recommendation result is triggered.

[0030] The three-layer structure is connected by a unified data channel, realizing a complete closed-loop path for data from sensory input, intelligent modeling, recommendation output to behavioral feedback.

[0031] This method uses a dynamic demand prediction engine based on graph neural networks and time series modeling algorithms to integrate user behavior, environmental context, and historical transaction data to construct a real-time updated demand probability field, which drives the generation of personalized recommendation strategies.

[0032] This method supports deep fusion processing of multimodal data sources, including more than 11 categories such as biometric information, spatial behavior trajectory, historical transaction records, member tags, voice signals, and facial expression status. It constructs dynamic user profiles and scene perception maps through unified encoding and context modeling. The method also extracts features and models the context of the data through a multimodal deep fusion mechanism to build high-precision dynamic user profiles and scene perception models.

[0033] Combination Figures 1 to 3 As shown, the specific implementation process of this method is as follows: like Figure 1 The diagram illustrates the three-layer technical architecture of this method in practical applications and the data interaction relationships between them: Data perception layer: This includes facial recognition terminals, IoT sensors, CRM system interfaces, and transaction POS data. It is mainly responsible for collecting raw data related to users and the environment.

[0034] Intelligent Decision Layer (Core Layer): Based on a dynamic image engine, a large demand prediction model, and a recommendation strategy generator, it realizes intelligent modeling and inference computation from perceived data to recommendation strategies.

[0035] The execution feedback layer includes multiple execution exits such as staff PAD push notifications, electronic price tag linkage, and self-service terminal guidance, responsible for quickly implementing strategy results. This layer also collects actual user behavior data and feeds it back to the perception layer, enabling the system to continuously learn and optimize.

[0036] 1. Deploy various sensing devices to form a data sensing layer, including the following key components: (1) A biometric terminal, including: Multispectral camera: It integrates visible light, infrared light and depth information to achieve accurate face recognition in complex environments and enhances robustness against occlusion. Voiceprint acquisition array: It adopts an anti-noise pickup design, which can still stably acquire voiceprint features in noisy environments, and support customer identity verification.

[0037] (2) Internet of Things (IoT) sensor networks, including: Shelf pressure sensor: With a measurement accuracy of 0.1g, it can sense the picking and placement of goods in real time and help judge customer dwell behavior; UWB positioning base station: Based on the TDOA (Time Difference of Arrival) algorithm, it achieves high-precision customer positioning and is used for path tracking and movement modeling.

[0038] (3) CRM system: The OAuth 2.0 authentication mechanism is used to connect to the enterprise membership management system to achieve customer identity synchronization and retrieval of historical behavior; (4) POS and inventory system data bus: Provides real-time data streams such as product inventory and sales records to support recommendation and replenishment logic calls.

[0039] 2. The technical process of the intelligent decision-making layer (core layer) is as follows: Figure 2 As shown.

[0040] The intelligent decision-making layer, as the core modeling and reasoning module, mainly includes the following processes: (1) Raw data input: Structured information from the data perception layer (such as user profiles, location trajectories, voice commands, product status, etc.) serves as the input basis.

[0041] (2) Feature extraction module: Performs multi-dimensional feature extraction on the original input, including numerical normalization, time-series slicing, context label completion, etc., to generate a unified vector representation.

[0042] (3) Dynamic fusion module: Performs fusion operation on multi-source heterogeneous data, adopts Transformer structure or self-attention mechanism, constructs context dependency matrix, and improves the model’s ability to perceive the intrinsic relationship between different input dimensions.

[0043] (4) Real-time streaming data enters the "demand forecasting model", such as time series analysis network, graph neural network (GNN) or ST-GNN module, to predict users' current and near-term potential demand; (5) Static image data (such as shelf status, customer facial status, etc.) are entered into the “Long-term Evolution of Profile” module to capture customer behavior habits and trend information; (6) Strategy optimization module: Combine the output results of each path, use reinforcement learning policy network or multi-objective optimization algorithm to output the optimal action suggestion (such as recommended products, navigation path, speech generation, etc.).

[0044] (7) Execution instruction generation: Form structured execution instructions and push them to various terminals (robots, shelf screens, IoT devices, etc.) in the interactive execution layer to reach end users.

[0045] 3. The execution feedback layer is a multi-channel adaptive outreach system that transforms recommendation strategies into actual marketing actions through smart terminal devices, while simultaneously collecting user feedback data in real time to complete closed-loop optimization.

[0046] The execution feedback layer mainly consists of the following three components: (1) PAD push for store staff: The system pushes customer identity, profile and personalized recommendation content to store staff mobile terminals (PAD) in real time to assist sales guides in targeted recommendations, door-to-door greetings and proactive interventions, thereby improving the efficiency and accuracy of manual intervention.

[0047] (2) Electronic price tag update: The system can control the electronic price tag information of the target product through the Internet of Things bus, and dynamically modify the price, slogan and remaining quantity prompt in scenarios such as inventory changes, promotion launch, and recommendation priority, so as to realize shelf-level linkage response.

[0048] (3) Self-service terminal guidance: Customers can obtain personalized recommendation lists, navigation paths and combination discount prompts through self-service terminals in the store (such as touch screens and smart shelf panels) to achieve a closed loop of self-service.

[0049] The following is the complete customer entry process and its corresponding technical implementation: Step 1: Identity recognition and demand prediction.

[0050] When a customer enters the store, multispectral cameras deployed at the entrance immediately activate to capture facial images. These cameras operate simultaneously in the visible, infrared, and depth bands, possessing strong light suppression and low light enhancement capabilities to ensure high-quality images are acquired even under complex lighting conditions such as backlighting and dim lighting. This also triggers the service robot (based on the ROS system) to wake up. The robot collects the customer's facial information through a multi-camera array (RGB + depth) and calls a locally deployed facial recognition API (based on ArcFace or FaceNet algorithms) for authentication. Upon successful recognition, the system connects to the CRM database (MySQL / Redis cache) via Redis caching to match member information. If the customer is identified as a registered existing customer, the system automatically generates a customized voice greeting, which is then broadcast in real-time using a speech synthesis model (Tacotron2 + WaveRNN), enhancing user engagement and friendliness.

[0051] The background recommendation engine (TensorFlow Serving) analyzes data in real time across three dimensions: The time series analysis module calls the Prophet model to perform time series modeling on customers' past orders and predict users' cyclical consumption trends.

[0052] Association rule mining (Apriori algorithm) combines customer information to extract associated products from the product knowledge graph constructed from the Neo4j graph database.

[0053] The real-time inventory system queries the product status in the ERP system in real time via REST API to ensure sufficient recommended products and effective promotional activities.

[0054] If this is a first-time customer, the system will guide them to register as a member via a WeChat mini-program. The WeChat mini-program API generates a dynamic QR code, and the form information is automatically entered into the member's profile using OCR recognition (based on PaddleOCR). The registration form includes: Date picker: Allows users to enter their birth date for subsequent recommendation strategy optimization.

[0055] Preference selector: Allows users to select product categories, price ranges, service types, allergens, and other personalized preferences, and the system builds recommendation filtering rules based on these.

[0056] The facial recognition database encryption system (using AES-256 encryption) encrypts and encodes the facial features of family members to construct a parent-child relationship map, which is used for product recommendation and service recognition in subsequent scenarios.

[0057] Step 2: Confirm and proceed to the target shelf.

[0058] Recommendations are displayed on an Android interactive screen. Customers can confirm or modify the recommendation list via the touchscreen. The robot activates its SLAM navigation system (LiDAR + IMU + UWB fusion positioning), uses an improved A* algorithm for path planning, and relies on the YOLOv5 real-time detection model for dynamic obstacle avoidance.

[0059] Upon reaching the target shelf, AR tag recognition (ARKit) assists in precise positioning, and the robot confirms the product's location through dual-mode positioning (QR code + RFID). If a trial pack is required, the nearest smart sampler is activated by calling the IoT device management system (MQTT protocol) to prepare the sample.

[0060] In addition, customer movement within the store is monitored in real time by an ultra-wideband positioning system, which, combined with multi-view smart cameras deployed on the ceiling, forms a three-dimensional positioning network with centimeter-level accuracy. When a customer lingers in front of a shelf for more than 3 seconds, and the computer vision system detects that their gaze angle forms a predetermined angle with the product display angle, the system automatically marks it as a high-intent behavior. This behavioral data is combined with weight change sensors from the smart shelves and product handling recognition results to construct a complete customer movement map. All sensor data is synchronized at the millisecond level through a precise time protocol, ensuring the spatiotemporal consistency of behavioral analysis.

[0061] Step 3: Adaptive scene processing.

[0062] During navigation, the system possesses the ability to automatically adapt to specific scenarios, ensuring customers receive the best service experience under varying needs. The system can automatically switch scene modes based on environmental perception, customer characteristics, and behavioral signals, achieving a more personalized and differentiated intelligent response, including but not limited to: Environmental Change Response: When the system detects a significant change in the environment (such as dim lighting or excessive noise), it will automatically switch to a more suitable interaction mode. For example, in noisy environments, it will reduce the frequency of voice announcements and increase the proportion of on-screen text prompts; in bright / low-light environments, it will automatically adjust the display brightness to ensure visibility.

[0063] User behavior state adaptation: Based on behavioral characteristics such as customer dwell time, changes in walking speed, and browsing frequency, the system can intelligently identify whether a user may be in a hesitant, fatigued, or highly attentive state, and recommend rest areas, popular products, and auxiliary instructions in a timely manner to improve thoughtfulness and smoothness of the experience.

[0064] Multi-user identification and collaborative recommendation: When multiple customers are detected interacting at the same time (such as multiple people viewing the terminal or browsing the same area at the same time), the system can dynamically coordinate multiple user profiles to provide combined recommendation schemes or synchronously adjust the interactive content to avoid information conflicts and optimize service paths.

[0065] Space guidance and prompts service: Based on the store layout and the customer's current location, the system can intelligently recommend the nearest service facilities or special areas, such as rest areas, self-checkout counters, trial areas, and information desks, enhancing the customer's sense of exploration and convenience.

[0066] Step 4: Feedback data collection (before leaving the store).

[0067] Before customers leave the store, the system automatically performs feedback data collection tasks to generate high-quality sample data required for closed-loop verification and model training.

[0068] First, the system automatically compares the customer's final actual purchase with the recommendation list generated by the intelligent decision-making layer to calculate the recommendation hit rate, which is used to measure the accuracy of the recommendation strategy.

[0069] Secondly, the facial recognition module deployed in the exit area will capture the facial expression features of customers, identify their subjective emotional state when leaving the store (such as satisfaction, confusion, fatigue, etc.), and combine it with voice tone analysis to judge potential satisfaction and interaction fluency.

[0070] In addition, the system guides users to provide proactive feedback through methods such as scanning QR codes, rating on self-service terminal pages, and voice inquiries, supplementing their subjective feelings and suggestions, and building an explicit feedback data layer.

[0071] All the feedback data described above undergoes multiple de-identification and differential privacy algorithms before storage to ensure complete anonymization of individual identity information, retaining only structured feature vectors used for model optimization. The feedback data is then synchronously fed into the model training module, supporting the policy engine for incremental learning, parameter updates, and iterative personalized recommendation path iterations.

[0072] Step 5: System self-evolution (ongoing).

[0073] This system possesses continuous learning and self-optimization capabilities. Through a federated learning architecture, it enables online evolution of store-level recommendation models. Each store deploys a lightweight model training engine locally, performing edge computing on customer behavior data (such as browsing paths, facial expression changes, and purchasing preferences), uploading only encrypted model parameters to ensure data privacy and security. All uploaded parameters are aggregated to a central server, with the system performing incremental updates hourly and full model retraining weekly. For new customers, the system employs a cold-start strategy combining knowledge graph inference and best-selling product recommendations to ensure reasonable suggestions are provided even without historical data. The system also integrates a log analysis system (such as ELK Stack) to record data throughout the service process. All data is uploaded to the cloud in real time, and the federated learning system regularly updates the recommendation models for each store. When a customer leaves the store, the intelligent checkout system (RFID + computer vision) automatically verifies the goods, automatically labels the purchase behavior, updates the user profile tags, and the recommendation engine pre-generates the next service strategy.

[0074] The proposed dynamic 3D recommendation engine integrates three dimensions—user consumption patterns, inventory replenishment cycles, and current scenario demands—to achieve a dynamically evolving multi-objective recommendation strategy. It possesses the following technical advantages: Strong semantic modeling capabilities: The recommendation strategy integrates user consumption frequency information, inventory replenishment cycle (ERP real-time data), and scenario demand tags. Through collaborative modeling of graph neural networks (GNN) and large language models (LLM), it achieves a unified expression of structured behavioral graphs and semantic intent.

[0075] Significantly improved recommendation accuracy: The system incorporates graph neural networks and a large-scale semantic engine. Actual testing shows that the accuracy of recommended products reaches 89%, significantly higher than the average level of the offline retail industry (approximately 60%). The effect is particularly outstanding in scenarios such as new product guidance and supplementary product recommendations.

[0076] Reliable and Sensitive Replenishment Forecasting: The replenishment alert module integrates consumption rhythm forecasting, activity cycle projection, and product consumption cycle analysis, employing an improved time series regression algorithm to predict replenishment timing. The error is controlled within ±3 days, significantly improving inventory turnover efficiency and operational responsiveness compared to the ±14-day accuracy of traditional manual judgment.

[0077] This invention also provides an intelligent recommendation and precision marketing system based on a large model, comprising: Data perception layer: This includes facial recognition terminals, IoT sensors, CRM system interfaces, and transaction POS data. It is mainly responsible for collecting raw data related to users and the environment.

[0078] Intelligent Decision Layer: Based on a dynamic image engine, a large demand prediction model, and a recommendation strategy generator, it realizes intelligent modeling and inference computation from perceived data to recommendation strategies.

[0079] Execution Feedback Layer: This layer includes multiple execution exits such as staff PAD push notifications, electronic price tag linkage, and self-service terminal guidance, responsible for quickly implementing strategy results. Simultaneously, this layer is also responsible for collecting actual user behavior data and feeding it back to the perception layer, enabling the system to continuously learn and optimize itself. This system achieves intelligent recommendation and precision marketing through the large-model-based intelligent recommendation and precision marketing method described in the above embodiments.

[0080] The data perception layer is composed of various sensing devices deployed in the system, and includes the following key components: (1) Biometric terminal: Multispectral camera: It integrates visible light, infrared light and depth information to achieve accurate face recognition in complex environments and enhances robustness against occlusion. Voiceprint acquisition array: Adopting an anti-noise pickup design, it can still stably acquire voiceprint features in noisy environments, supporting auxiliary verification of customer identity. (2) Internet of Things (IoT) sensor network: Shelf pressure sensor: With a measurement accuracy of 0.1g, it can sense the picking and placement of goods in real time and help judge customer dwell behavior; UWB positioning base station: Based on the TDOA (Time Difference of Arrival) algorithm, it achieves high-precision customer positioning and is used for path tracking and movement modeling.

[0081] (3) CRM system: The OAuth 2.0 authentication mechanism is used to connect to the enterprise membership management system to realize the synchronization of customer identity and retrieval of historical behavior.

[0082] (4) POS and inventory system data bus: Provides real-time data streams such as product inventory and sales records to support recommendation and replenishment logic calls.

[0083] The intelligent decision-making layer, as the core modeling and reasoning module of the system, mainly includes the following processes: (1) Raw data input: Structured information from the data perception layer (such as user profiles, location trajectories, voice commands, product status, etc.) serves as the input basis.

[0084] (2) Feature extraction module: Performs multi-dimensional feature extraction on the original input, including numerical normalization, time-series slicing, context label completion, etc., to generate a unified vector representation.

[0085] (3) Dynamic fusion module: Performs fusion operation on multi-source heterogeneous data, adopts Transformer structure or self-attention mechanism, constructs context dependency matrix, and improves the model’s ability to perceive the intrinsic relationship between different input dimensions.

[0086] (4) Real-time streaming data enters the "demand forecasting model", such as time series analysis network, graph neural network (GNN) or ST-GNN module, to predict users' current and near-term potential demand; (5) Static image data (such as shelf status, customer facial status, etc.) are entered into the “Long-term Evolution of Profile” module to capture customer behavior habits and trend information; (6) Strategy optimization module: Combine the output results of each path, use reinforcement learning policy network or multi-objective optimization algorithm to output the optimal action suggestion (such as recommended products, navigation path, speech generation, etc.).

[0087] (7) Execution instruction generation: Form structured execution instructions and push them to various terminals (robots, shelf screens, IoT devices, etc.) in the interactive execution layer to reach end users.

[0088] The execution feedback layer is a multi-channel adaptive outreach system that transforms recommendation strategies into actual marketing actions through smart terminal devices, while simultaneously collecting user feedback data in real time to complete closed-loop optimization.

[0089] The execution feedback layer mainly consists of the following three components: (1) PAD push for store staff: The system pushes customer identity, profile and personalized recommendation content to store staff mobile terminals (PAD) in real time to assist sales guides in targeted recommendations, door-to-door greetings and proactive interventions, thereby improving the efficiency and accuracy of manual intervention.

[0090] (2) Electronic price tag update: The system can control the electronic price tag information of the target product through the Internet of Things bus, and dynamically modify the price, slogan and remaining quantity prompt in scenarios such as inventory changes, promotion launch, and recommendation priority, so as to realize shelf-level linkage response.

[0091] (3) Self-service terminal guidance: Customers can obtain personalized recommendation lists, navigation paths and combination discount prompts through self-service terminals in the store (such as touch screens and smart shelf panels) to achieve a closed loop of self-service.

[0092] The entire process of intelligent recommendations implemented by this system when a customer enters the store is as follows: Step 1: Identity recognition and demand prediction.

[0093] When a customer enters the store, multispectral cameras deployed at the entrance immediately activate to capture facial images. These cameras operate simultaneously in the visible, infrared, and depth bands, possessing strong light suppression and low light enhancement capabilities to ensure high-quality images are acquired even under complex lighting conditions such as backlighting and dim lighting. This also triggers the service robot (based on the ROS system) to wake up. The robot collects the customer's facial information through a multi-camera array (RGB + depth) and calls a locally deployed facial recognition API (based on ArcFace or FaceNet algorithms) for authentication. Upon successful recognition, the system connects to the CRM database (MySQL / Redis cache) via Redis caching to match member information. If the customer is identified as a registered existing customer, the system automatically generates a customized voice greeting, which is then broadcast in real-time using a speech synthesis model (Tacotron2 + WaveRNN), enhancing user engagement and friendliness.

[0094] The background recommendation engine (TensorFlow Serving) analyzes data in real time across three dimensions: The time series analysis module calls the Prophet model to perform time series modeling on customers' past orders and predict users' cyclical consumption trends.

[0095] Association rule mining (Apriori algorithm) combines customer information to extract associated products from the product knowledge graph constructed from the Neo4j graph database.

[0096] The real-time inventory system queries the product status in the ERP system in real time via REST API to ensure sufficient recommended products and effective promotional activities.

[0097] If this is a first-time customer, the system will guide them to register as a member via a WeChat mini-program. The WeChat mini-program API generates a dynamic QR code, and the form information is automatically entered into the member's profile using OCR recognition (based on PaddleOCR). The registration form includes: Date picker: Allows users to enter their birth date for subsequent recommendation strategy optimization.

[0098] Preference selector: Allows users to select product categories, price ranges, service types, allergens, and other personalized preferences, and the system builds recommendation filtering rules based on these.

[0099] The facial recognition database encryption system (using AES-256 encryption) encrypts and encodes the facial features of family members to construct a parent-child relationship map, which is used for product recommendation and service recognition in subsequent scenarios.

[0100] Step 2: Confirm and proceed to the target shelf.

[0101] Recommendations are displayed on an Android interactive screen. Customers can confirm or modify the recommendation list via the touchscreen. The robot activates its SLAM navigation system (LiDAR + IMU + UWB fusion positioning), uses an improved A* algorithm for path planning, and relies on the YOLOv5 real-time detection model for dynamic obstacle avoidance.

[0102] Upon reaching the target shelf, AR tag recognition (ARKit) assists in precise positioning, and the robot confirms the product's location through dual-mode positioning (QR code + RFID). If a trial pack is required, the nearest smart sampler is activated by calling the IoT device management system (MQTT protocol) to prepare the sample.

[0103] In addition, customer movement within the store is monitored in real time by an ultra-wideband positioning system, which, combined with multi-view smart cameras deployed on the ceiling, forms a three-dimensional positioning network with centimeter-level accuracy. When a customer lingers in front of a shelf for more than 3 seconds, and the computer vision system detects that their gaze angle forms a predetermined angle with the product display angle, the system automatically marks it as a high-intent behavior. This behavioral data is combined with weight change sensors from the smart shelves and product handling recognition results to construct a complete customer movement map. All sensor data is synchronized at the millisecond level through a precise time protocol, ensuring the spatiotemporal consistency of behavioral analysis.

[0104] Step 3: Adaptive scene processing.

[0105] During navigation, the system possesses the ability to automatically adapt to specific scenarios, ensuring customers receive the best service experience under varying needs. The system can automatically switch scene modes based on environmental perception, customer characteristics, and behavioral signals, achieving a more personalized and differentiated intelligent response, including but not limited to: Environmental Change Response: When the system detects a significant change in the environment (such as dim lighting or excessive noise), it will automatically switch to a more suitable interaction mode. For example, in noisy environments, it will reduce the frequency of voice announcements and increase the proportion of on-screen text prompts; in bright / low-light environments, it will automatically adjust the display brightness to ensure visibility.

[0106] User behavior state adaptation: Based on behavioral characteristics such as customer dwell time, changes in walking speed, and browsing frequency, the system can intelligently identify whether a user may be in a hesitant, fatigued, or highly attentive state, and recommend rest areas, popular products, and auxiliary instructions in a timely manner to improve thoughtfulness and smoothness of the experience.

[0107] Multi-user identification and collaborative recommendation: When multiple customers are detected interacting at the same time (such as multiple people viewing the terminal or browsing the same area at the same time), the system can dynamically coordinate multiple user profiles to provide combined recommendation schemes or synchronously adjust the interactive content to avoid information conflicts and optimize service paths.

[0108] Space guidance and prompts service: Based on the store layout and the customer's current location, the system can intelligently recommend the nearest service facilities or special areas, such as rest areas, self-checkout counters, trial areas, and information desks, enhancing the customer's sense of exploration and convenience.

[0109] Step 4: Feedback data collection (before leaving the store).

[0110] Before customers leave the store, the system automatically performs feedback data collection tasks to generate high-quality sample data required for closed-loop verification and model training.

[0111] First, the system automatically compares the customer's final actual purchase with the recommendation list generated by the intelligent decision-making layer to calculate the recommendation hit rate, which is used to measure the accuracy of the recommendation strategy.

[0112] Secondly, the facial recognition module deployed in the exit area will capture the facial expression features of customers, identify their subjective emotional state when leaving the store (such as satisfaction, confusion, fatigue, etc.), and combine it with voice tone analysis to judge potential satisfaction and interaction fluency.

[0113] In addition, the system guides users to provide proactive feedback through methods such as scanning QR codes, rating on self-service terminal pages, and voice inquiries, supplementing their subjective feelings and suggestions, and building an explicit feedback data layer.

[0114] All the feedback data described above undergoes multiple de-identification and differential privacy algorithms before storage to ensure complete anonymization of individual identity information, retaining only structured feature vectors used for model optimization. The feedback data is then synchronously fed into the model training module, supporting the policy engine for incremental learning, parameter updates, and iterative personalized recommendation path iterations.

[0115] Step 5: System self-evolution (ongoing).

[0116] This system possesses continuous learning and self-optimization capabilities. Through a federated learning architecture, it enables online evolution of store-level recommendation models. Each store deploys a lightweight model training engine locally, performing edge computing on customer behavior data (such as browsing paths, facial expression changes, and purchasing preferences), uploading only encrypted model parameters to ensure data privacy and security. All uploaded parameters are aggregated to a central server, with the system performing incremental updates hourly and full model retraining weekly. For new customers, the system employs a cold-start strategy combining knowledge graph inference and best-selling product recommendations to ensure reasonable suggestions are provided even without historical data. The system also integrates a log analysis system (such as ELK Stack) to record data throughout the service process. All data is uploaded to the cloud in real time, and the federated learning system regularly updates the recommendation models for each store. When a customer leaves the store, the intelligent checkout system (RFID + computer vision) automatically verifies the goods, automatically labels the purchase behavior, updates the user profile tags, and the recommendation engine pre-generates the next service strategy.

[0117] This system targets offline retail scenarios, enabling intelligent recommendations with high requirements for accurate customer identification, real-time demand insight, efficient service interaction, and privacy protection. Leveraging core technologies such as large language models, multimodal fusion recognition, knowledge graph reasoning, robot-guided navigation, and federated learning, the system constructs a complete service loop of "entry recognition—demand prediction—dynamic recommendation—path guidance—feedback learning," enhancing the personalization of in-store services, the efficiency of sales staff responses, and customer loyalty. It provides offline retail scenarios with intelligent, automated, and sustainably optimized precision marketing capabilities. Simultaneously, it addresses the core pain points of traditional marketing, namely "blind recommendation" and "passive service," specifically including: Solving the retail dilemma of "abundant data but weak insights": Traditional retailers possess over 70% of user-related data, but their data utilization rate is less than 15%, hindering personalized services. This invention's system transforms discrete data (such as purchase records, movement patterns, and facial expression changes) into a continuous demand probability field and constructs a three-dimensional correlation matrix of user-product-scenario based on a spatiotemporal graphical neural network (ST-GNN), enabling deep insights and real-time responses to customers' potential needs.

[0118] Enhancing the breadth and depth of service capabilities: In traditional offline stores, when one employee needs to serve 3 to 5 customers simultaneously, the accuracy of recommendations often drops by more than 60%, severely impacting the user experience. This invention's system supports automatic association of member identities for up to 100 concurrent users (100 people / second) using facial recognition technology, while simultaneously using a large language model to generate personalized recommendation scripts in real time (response latency <0.5 seconds), ensuring that the depth of recommendations is not weakened while improving the breadth of service.

[0119] This invention also provides an intelligent recommendation precision marketing device based on a large model, comprising: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is used to call the machine-readable program to implement the intelligent recommendation and precision marketing method based on a large model as described in the above embodiments.

[0120] This invention also provides a computer-readable medium storing computer instructions. When executed by a processor, the computer instructions cause the processor to perform the intelligent recommendation and precision marketing method based on a large model as described in the above embodiments. Specifically, a system or apparatus equipped with a storage medium storing software program code that implements the functions of any of the embodiments described above can be provided, and the computer (or CPU or MPU) of the system or apparatus can read and execute the program code stored in the storage medium.

[0121] In this case, the program code read from the storage medium can itself implement the function of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute part of the present invention.

[0122] Storage media embodiments for providing program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer via a communication network.

[0123] Furthermore, it should be clear that not only can the program code read by the computer be executed, but also the operating system or other components operating on the computer can be instructed based on the program code to perform some or all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0124] Furthermore, it is understood that the program code read from the storage medium is written to the memory set in the expansion board inserted into the computer or to the memory set in the expansion unit connected to the computer. Then, based on the instructions of the program code, the CPU or other components installed on the expansion board or expansion unit execute some and all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0125] The present invention has been shown and described in detail above with reference to the accompanying drawings and preferred embodiments. However, the present invention is not limited to these disclosed embodiments. Based on the above embodiments, those skilled in the art will know that more embodiments of the present invention can be obtained by combining the code review methods in the different embodiments. These embodiments are also within the protection scope of the present invention.

Claims

1. A method for intelligent recommendation and precision marketing based on a large model, characterized in that, The implementation of this method includes: The data perception layer is used to collect user-related information, including facial recognition terminals, IoT sensors, CRM systems, and transaction POS data. The intelligent decision-making layer is used to fuse and model the perceived data and generate recommendation strategies, and includes a dynamic image engine, a large demand prediction model, and a recommendation strategy generator. The execution feedback layer is used to deliver recommendation results to users through multiple channels and collect feedback data. It includes a staff PAD push module, an electronic price tag update module, and a self-service terminal guidance module.

2. The intelligent recommendation and precision marketing method based on a large model according to claim 1, characterized in that, By using a dynamic demand prediction engine based on graph neural networks and time series modeling algorithms, and integrating user behavior, environmental context, and historical transaction data, a real-time updated demand probability field is constructed to drive the generation of personalized recommendation strategies.

3. The intelligent recommendation and precision marketing method based on a large model according to claim 1, characterized in that, It supports deep fusion processing of multimodal data sources, including biometric information, spatial behavior trajectories, historical transaction records, member tags, voice signals, and facial expression states. Dynamic user profiles and scene perception maps are constructed through unified encoding and context modeling. Feature extraction and context modeling are performed on the data through a multimodal deep fusion mechanism to build high-precision dynamic user profiles and scene perception models.

4. The intelligent recommendation and precision marketing method based on a large model according to claim 1, characterized in that, For the data perception layer The face recognition terminal includes: Multispectral camera: used to fuse visible light, infrared light and depth information to achieve accurate face recognition in complex environments and enhance occlusion robustness; Voiceprint acquisition array: It adopts an anti-noise pickup design, which can still stably acquire voiceprint features in noisy environments, and support customer identity verification. The IoT sensor includes: Shelf pressure sensors: used to sense the picking and placement of goods in real time, helping to determine customer dwell behavior; UWB positioning base station: Based on the TDOA algorithm, it achieves high-precision customer positioning and is used for path tracking and movement modeling; The CRM system, Integrate with the enterprise membership management system to synchronize customer identities and retrieve historical behavior data; The transaction POS data, Provides real-time data streams including product inventory and sales records to support recommendation and replenishment logic calls.

5. The intelligent recommendation and precision marketing method based on a large model according to claim 1, characterized in that, The demand forecasting model is based on graph neural networks and time series forecasting methods, and constructs a three-dimensional demand probability matrix of users, products and scenarios. The recommendation strategy generator generates personalized recommendation scripts that match the user profile by calling a large language model; The intelligent decision-making layer includes the following implementation process: (1) Raw data input: The structured information from the data perception layer is used as the input basis; (2) Feature extraction: Perform multi-dimensional feature extraction on the original input, including numerical normalization, time-series slicing, context label completion, and generate a unified vector representation; (3) Dynamic fusion: Perform fusion operation on multi-source heterogeneous data, use Transformer structure or self-attention mechanism to construct context dependency matrix, and improve the model’s ability to perceive the intrinsic relationship between different input dimensions; (4) Real-time streaming data enters the "demand forecasting model", including time series analysis network, graph neural network or ST-GNN module, to predict users' current and near-term potential demand; (5) Static image data is entered into the "Long-term Evolution of Profile" module to capture customer behavior habits and trend information; (6) Strategy optimization: Combine the output results of each path, and use reinforcement learning policy network or multi-objective optimization algorithm to output the optimal action suggestion; (7) Execution instruction generation: Form structured execution instructions and push them to each terminal of the interactive execution layer to reach the end user.

6. The intelligent recommendation and precision marketing method based on a large model according to claim 1, characterized in that, For the execution feedback layer, The store clerk PAD push module synchronizes user profiles and recommended products to the service personnel's terminals, assisting them in manual guidance and intervention operations. The electronic price tag update module dynamically controls the content displayed on the shelf price tags based on inventory information and recommendation priority, enabling promotional synchronization and inventory reminders. The self-service terminal guidance module allows users to view the recommendation logic, obtain the navigation path, and complete the QR code confirmation on the terminal interface, thus realizing a closed loop of self-service. The feedback data includes product response behavior, customer emotional characteristics, and explicit rating results, which are then anonymized before being used in the training of the federated learning model.

7. The intelligent recommendation and precision marketing method based on a large model according to claim 1, characterized in that, The specific process for achieving intelligent recommendation and precise marketing when customers enter the store includes: Step 1: Identity Recognition and Demand Prediction: When a customer enters the store, the multispectral camera deployed at the entrance immediately activates to capture facial images, simultaneously waking up the service robot. The robot collects the customer's facial information through the multi-camera array and calls the locally deployed facial recognition API for identity verification. Upon successful recognition, it connects to the CRM database via Redis cache to match member information. If the customer is identified as a registered existing customer, a customized voice greeting is automatically generated and broadcast in real time. The backend recommendation engine analyzes data from three dimensions in real time: The time series analysis module calls the Prophet model to perform time series modeling on customers' past orders and predict users' periodic consumption trends; Association rule mining combines customer information to extract related products from the product knowledge graph constructed from the Neo4j graph database; The real-time inventory system queries the product status in the ERP system in real time via REST API to ensure sufficient recommended products and effective promotional activities; If this is a first-time customer, the system will guide them through a membership registration process. The form information will be automatically entered into the customer's database after being recognized by OCR. The registration form includes: Date picker: Used by users to enter their date of birth, which is then used to optimize subsequent recommendation strategies; Preference selector: Allows users to select personalized preferences, including product categories, price ranges, service types, and allergens of interest, and the system builds recommendation filtering rules based on these preferences; The facial recognition database encryption system encrypts and encodes the facial features of family members to construct a parent-child relationship map, which can be used for product recommendation and service recognition in subsequent scenarios. Step 2: Confirm and proceed to the target shelf: Recommendations are displayed on an interactive screen; customers confirm or modify the recommendation list via a touchscreen, and the robot activates the SLAM navigation system; upon reaching the target shelf, AR marker recognition assists in precise positioning, and the robot confirms the product location through dual-mode positioning; if a trial pack is needed, the nearest smart sampler is activated by calling the IoT device management system to prepare the sample; Meanwhile, the customer's movement trajectory in the store is monitored in real time by an ultra-wideband positioning system, which, together with multi-view smart cameras deployed on the ceiling, forms a three-dimensional positioning network with centimeter-level accuracy. When a customer stays in front of a shelf for more than a set time, and the computer vision system detects that the angle of their line of sight forms a set angle with the angle of the merchandise display, the system automatically marks it as a high-intent behavior. These behavioral data are combined with the weight change sensors from the smart shelves and the results of merchandise picking and placing actions to construct a complete customer movement map. Step 3: Scene Adaptive Processing During navigation guidance, the system has automatic adaptive capabilities, including: Environmental change response: When the system detects a significant change in the field environment, it will automatically switch to a more suitable interaction mode; User behavior state adaptation: Based on customer behavior characteristics, including dwell time, walking speed changes, and browsing frequency, intelligently identify whether the user is in a hesitant, fatigued, or highly attentive state, and recommend rest areas, popular products, and auxiliary instructions; Multi-user identification and collaborative recommendation: When multiple customers are detected interacting at the same time, the multi-user profiles are dynamically coordinated to provide combined recommendation solutions or adjust the interaction content synchronously to avoid information conflicts and optimize the service path; Space guidance and prompt service: Based on the store layout and the customer's current location, intelligently recommend the nearest service facilities or featured areas; Step 4: Feedback Data Collection Before customers leave the store, the system automatically performs a feedback data collection task, generating high-quality sample data required for closed-loop validation and model training; including: First, the actual products purchased by customers are automatically compared with the recommended list generated by the intelligent decision-making layer to calculate the recommendation hit rate, which is used to measure the accuracy of the recommendation strategy. Secondly, the facial recognition module deployed in the exit area captures the facial expression features of customers, identifies their subjective emotional state when leaving the store, and combines it with voice tone analysis to judge potential satisfaction and interaction fluency. In addition, users are encouraged to provide proactive feedback through QR code scanning, self-service terminal page rating, and voice inquiry, supplementing their subjective feelings and suggestions, and building an explicit feedback data layer. All feedback data undergoes multiple desensitization and differential privacy algorithms before storage to ensure complete anonymization of individual identity information, retaining only structured feature vectors used for model optimization; the feedback data will be synchronously fed into the model training module, supporting the policy engine for incremental learning, parameter updates, and personalized recommendation path iteration; Step 5: System Self-Evolution The system has the ability to continuously learn and self-optimize, and through a federated learning architecture, it enables the online evolution of the store-level recommendation model; Each store deploys a lightweight model training engine locally to perform edge computing on customer behavior data and only uploads encrypted model parameters; all uploaded parameters are aggregated to a central server, and the system performs incremental updates every hour and full model retraining every week; for new customers, a cold start strategy combining knowledge graph reasoning and best-selling product recommendations is adopted. An integrated log analysis system enables full-process data recording of services. All data will be uploaded to the cloud in real time, and the federated learning system will regularly update the recommendation models of each store. When a customer leaves the store, the intelligent checkout system automatically verifies the goods, automatically marks the purchase, updates the user profile tags, and the recommendation engine pre-generates the next service strategy.

8. A large-scale intelligent recommendation precision marketing system, characterized in that, include: Data perception layer: including facial recognition terminals, IoT sensors, CRM system interfaces and transaction POS data, responsible for collecting raw data related to users and the environment; Intelligent Decision Layer: Based on a dynamic image engine, a large demand prediction model, and a recommendation strategy generator, it realizes intelligent modeling and inference computation from perceived data to recommendation strategies; Execution Feedback Layer: This layer includes multiple execution exits such as staff PAD push notifications, electronic price tag linkage, and self-service terminal guidance, and is responsible for quickly implementing strategy results. At the same time, this layer is also responsible for collecting actual user behavior data and feeding it back to the perception layer to enable the system to continuously learn and optimize itself. The system achieves intelligent recommendation and precise marketing through the method described in any one of claims 1 to 7.

9. A smart recommendation precision marketing device based on a large model, characterized in that, include: At least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is configured to invoke the machine-readable program to implement the method according to any one of claims 1 to 7.

10. A computer-readable medium, characterized in that, The computer-readable medium stores computer instructions that, when executed by a processor, enable the implementation of the method described in any one of claims 1 to 7.

Images