Healthy retail supply chain optimization method and device and computer equipment

Abstract

The invention relates to the technical field of online sales, and discloses a healthy retail supply chain optimization method and device and computer device.The method comprises the steps that multi-level semantic analysis is conducted on obtained text data through a natural language processing technology, and potential demand keywords of a user are extracted; constructing a user portrait based on the user potential demand keyword and the user behavior data; grouping the users based on a preset grouping standard and the user portraits in the health scene, screening corresponding required health products for different user groups, and optimizing and adjusting the required health products in combination with market trend data in the user behavior data to obtain an adjusted product selection pool; and calculating a matching degree between the user and a preset candidate social platform based on the user portrait and the product selection pool, and generating a social platform recommendation list based on the matching degree. According to the method, multi-source heterogeneous data are integrated, potential demands of users are mined by using a large model, a product selection pool is optimized, and accurate recommendation and personalized marketing are realized in combination with user portrait analysis.

Description

Technical Field

[0001] The present invention relates to the field of online sales technology, and in particular to a health retail supply chain optimization method, device and computer equipment. Background Art

[0002] In the traditional health retail supply chain, product selection relies on manual experience and lacks data support, resulting in blind product selection and difficulty in meeting user needs. Existing recommendation systems are mostly based on a single data source (such as historical orders), which cannot fully tap into users' potential needs and have low recommendation accuracy. Marketing strategies lack personalization, making it difficult to develop differentiated strategies based on the price sensitivity and preferences of different users, resulting in low conversion rates. Multi-source heterogeneous data on social platforms is difficult to effectively integrate and utilize, and the value of the data is not fully tapped. The above technical problems encountered in the health retail supply chain make it difficult to achieve accurate recommendations and personalized marketing for health products and social platforms. Summary of the Invention

[0003] In view of this, the present invention provides a health retail supply chain optimization method, apparatus and computer equipment to solve the problem of difficulty in achieving accurate recommendations of health products and social platforms.

[0004] In a first aspect, the present invention provides a method for optimizing a health retail supply chain, the method comprising:

[0005] Obtain text data and user behavior data for health retail scenarios;

[0006] Use natural language processing technology to conduct multi-level semantic analysis on text data and extract keywords of users' potential needs;

[0007] Build user profiles based on user potential demand keywords and user behavior data;

[0008] Users are grouped based on preset grouping criteria and user portraits in health scenarios, and corresponding health products are screened for different user groups. Furthermore, these products are optimized and adjusted based on market trend data from user behavior data to obtain an adjusted product pool.

[0009] Based on the user portrait and product selection pool, the matching degree between the user and the preset candidate social platforms is calculated, and a recommended list of social platforms is generated based on the matching degree.

[0010] The present invention provides a health retail supply chain optimization method, which obtains text data and user behavior data of health retail scenarios, innovatively integrates multi-source data such as text data and user behavior data, and provides a comprehensive data foundation for the optimization of the health retail supply chain. Natural language processing technology is introduced to perform in-depth semantic analysis of text data, breaking through the limitations of traditional methods and accurately extracting keywords of users' potential needs. Multi-dimensional user portraits are constructed in combination with user behavior data to achieve a more comprehensive and in-depth understanding of user needs. Users are scientifically grouped, and suitable health products are screened according to the characteristics of different groups to improve the accuracy of product selection. The product selection pool is dynamically adjusted in combination with market trend data to make the product selection more in line with the dynamic market needs. Based on the user portrait and the product selection pool, the matching degree between the user and the preset candidate social platform is calculated, and a personalized social platform recommendation list is generated based on the matching degree. This enables health products to accurately reach target users on appropriate social platforms, improve promotion efficiency, reduce marketing costs, and achieve accurate recommendation of social platforms, solving the problem that it is difficult to achieve accurate recommendation of health products and social platforms in the existing technology.

[0011] In an optional embodiment, natural language processing technology is used to perform multi-level semantic analysis on text data to extract keywords of potential user needs, including:

[0012] Natural language processing technology is used to perform sequence labeling, sentiment polarity analysis, and topic distribution extraction on text data, respectively, to identify user intent, health topic attitudes, and topic distribution;

[0013] Generate explicit demand keywords and implicit demand keywords based on user intention, attitude towards health topics, topic distribution and preset health knowledge graph.

[0014] The present invention provides a health retail supply chain optimization method, which can deeply explore users' potential needs by acquiring text data and user behavior data of health retail scenarios and performing multi-level semantic analysis using natural language processing technology. Compared with traditional methods, it is no longer limited to surface needs, and the extracted potential demand keywords are more in line with users' real thoughts. Sequence labeling is used to identify user intentions, sentiment polarity analysis is used to reveal attitudes towards health topics, and topic distribution is used to extract clear discussion topics, thereby achieving a three-dimensional deconstruction of text data and comprehensively capturing the explicit and implicit information expressed by users. Combined with the preset health knowledge graph, the semantic analysis results are mapped to the professional health field knowledge system, so that keyword extraction conforms to language logic and improves the accuracy and practicality of demand keywords. Through sentiment analysis and knowledge graph reasoning, it is possible to identify users' implicit but not directly expressed health needs (such as preferences for specific ingredients or potential health concerns).

[0015] In an optional embodiment, the user behavior data includes social platform short video data, historical order data, social platform interaction data, and external data;

[0016] Build user profiles based on user potential demand keywords and user behavior data, including:

[0017] Obtain user basic attribute characteristics and extract health demand characteristics based on user potential demand keywords;

[0018] Extract consumer demand characteristics based on social platform short video data, historical order data, social platform interaction data and external data;

[0019] Build user portraits based on basic attribute characteristics, health needs characteristics and consumption needs characteristics.

[0020] The present invention provides a health retail supply chain optimization method that integrates multi-source data such as short video data from social platforms, historical order data, social platform interaction data, and external data. Compared with a single data source, it can fully capture the user's behavioral performance in different scenarios. The integration of multi-source data avoids the omission of user characteristics and provides rich and complete data support for portrait construction. Combined with the user's potential demand keywords, the implicit needs mined from text data and behavioral data complement each other, focusing on both the user's surface behavior and in-depth insight into the inner needs, ensuring that the user portrait accurately covers all aspects of the user's characteristics. The basic attribute characteristics, health demand characteristics, and consumer demand characteristics are extracted separately, and the user is analyzed from different dimensions. The basic attribute characteristics clarify the user's basic background, the health demand characteristics focus on the user's health demands, and the consumer demand characteristics show the user's consumption behavior pattern. The fusion of multi-dimensional features makes the user portrait three-dimensional and full, and can deeply portray the user's image.

[0021] In an optional embodiment, the preset grouping criteria include health demand characteristic criteria, consumption behavior characteristic criteria, health status characteristic criteria, and spatiotemporal characteristic criteria;

[0022] Group users based on preset grouping criteria and user portraits in health scenarios, including:

[0023] Based on health demand characteristic standards, consumption behavior characteristic standards, health status characteristic standards, spatiotemporal characteristic standards and user portraits, a clustering algorithm is used to group users to obtain different user groups.

[0024] In an optional embodiment, the market trend data in the user behavior data is combined to optimize and adjust the health products in demand to obtain an adjusted product pool, including:

[0025] Extract trend feature vectors from market trend data, and input the trend feature vectors into a preset time series prediction model for training to obtain trend prediction results;

[0026] Adjust the selection weights of in-demand health products based on trend forecast results and preset selection decision rules;

[0027] Based on the adjusted weights, in-demand health products are dynamically updated and removed from the shelves to obtain an adjusted product selection pool.

[0028] The present invention provides a health retail supply chain optimization method that, based on health demand characteristic standards, can clearly identify different users' differentiated needs for health products, health information, and health management services. In the fiercely competitive health market, user segmentation strategies based on preset segmentation criteria and user profiles enable companies to gain a deeper understanding of user groups and market segments that competitors have yet to fully tap. By accurately targeting these potential markets, companies can develop differentiated market competition strategies and provide unique health products and services. Trend feature vectors are extracted from market trend data, encompassing multi-dimensional information such as industry policy changes, competitive product dynamics, and evolving consumer preferences. This data-driven approach is more objective and accurate. Trend feature vectors are trained using a preset time series prediction model to uncover temporal correlations and underlying patterns in market trend data. Product selection weights are adjusted based on trend prediction results and preset product selection decision rules. This combines the predicted data with the company's strategic goals and business strategy, avoiding blind and subjective product selection. Combined with user segmentation strategies based on preset segmentation criteria and user profiles, the product selection pool can be optimized based on the needs of different user groups and market trends.

[0029] In an optional embodiment, the matching degree between the user and the preset candidate social platforms is calculated based on the user portrait and the product selection pool, and a social platform recommendation list is generated based on the matching degree, including:

[0030] Use embedding technology to convert user profiles into dense vectors;

[0031] Extract corresponding platform features of preset candidate social platforms and represent them in vector form; platform features include platform content, platform products, and real-time performance;

[0032] Based on the product selection pool, a hybrid collaborative filtering algorithm is used to calculate multiple matching degrees between dense vectors and corresponding platform feature vectors;

[0033] Arrange multiple matching degrees in descending order, and generate a social platform recommendation list based on the matching degrees after descending order and the social platform operation strategy.

[0034] The present invention provides a health retail supply chain optimization method, which uses embedding technology to convert user portraits into dense vectors, effectively reducing data dimensions while retaining multi-dimensional feature information of users, achieving efficient compression and accurate expression of user features, and vectorizing features such as platform content, platform products, and real-time performance of preset candidate social platforms, comprehensively and quantitatively presenting the unique attributes of each social platform. Based on the product selection pool, a hybrid collaborative filtering algorithm is used, combining user historical behavior data and product information in the product selection pool, comprehensively considering the similarity between users and the adaptability of products to social platforms, avoiding the limitations of a single collaborative filtering algorithm, and improving the accuracy of calculating the matching degree between users and social platforms. After arranging the matching degrees in descending order, a recommendation list is generated in combination with the social platform operation strategy, which helps companies achieve their business goals while meeting the personalized needs of users.

[0035] In an optional embodiment, generating a social platform recommendation list based on the matching degrees and social platform operation strategies after sorting in descending order includes:

[0036] Obtain real-time user behavior on social platforms and dynamically adjust social platform recommendation lists based on real-time user behavior.

[0037] The present invention provides a health retail supply chain optimization method that captures user behavior on social platforms in real time and dynamically adjusts recommendation lists based on real-time behavior, making the recommended content more immediate and relevant, allowing users to experience the personalized and intelligent nature of the recommendation service.

[0038] In an optional embodiment, the health retail supply chain optimization method further includes:

[0039] Obtain users' historical consumption behavior and price sensitivity;

[0040] In the process of generating recommendation lists on social platforms, we use operations research and optimization theory to develop personalized coupon distribution strategies based on historical consumption behavior, price sensitivity, and real-time user behavior.

[0041] Use A / B testing algorithms to optimize personalized coupon distribution strategies.

[0042] This invention provides a health retail supply chain optimization method that uses operations research theory to formulate coupon strategies. This method can rationally set coupon values, usage thresholds, and expiration dates based on user spending habits and price preferences, enabling precision marketing and improving the efficiency of marketing resource utilization. Personalized coupon issuance strategies can meet users' individual needs, enhance user engagement with and participation in promotional activities, and improve conversion rates. This personalized coupon issuance strategy is optimized using an A / B testing algorithm. By conducting comparative experiments using different coupon schemes, collecting user feedback data, and analyzing the impact of different strategies on user purchasing behavior, the optimal coupon issuance strategy is identified.

[0043] In a second aspect, the present invention provides a health retail supply chain optimization device, the device comprising:

[0044] Data acquisition module, used to obtain text data and user behavior data of health retail scenarios;

[0045] Keyword extraction module, which uses natural language processing technology to perform multi-level semantic analysis on text data and extract keywords of users' potential needs;

[0046] A user portrait building module, used to build a user portrait based on the user's potential demand keywords and the user behavior data;

[0047] The user grouping and product selection pool adjustment module is used to group users based on preset grouping criteria and user portraits in health scenarios, screen out corresponding health products for different user groups, and optimize and adjust the health products in demand based on market trend data in user behavior data to obtain an adjusted product selection pool;

[0048] The recommendation list generation module is used to calculate the matching degree between users and preset candidate social platforms based on user portraits and product selection pools, and generate a social platform recommendation list based on the matching degree.

[0049] In a third aspect, the present invention provides a computer device comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the healthy retail supply chain optimization method of the first aspect or any corresponding embodiment thereof by executing the computer instructions.

[0050] In a fourth aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to enable a computer to execute the healthy retail supply chain optimization method of the first aspect or any corresponding embodiment thereof.

[0051] In a fifth aspect, the present invention provides a computer program product comprising computer instructions for causing a computer to execute the health retail supply chain optimization method of the first aspect or any corresponding embodiment thereof. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 is a flow chart of a method for optimizing a health retail supply chain according to an embodiment of the present invention;

[0054] Figure 2 is a flow chart of another health retail supply chain optimization method according to an embodiment of the present invention;

[0055] Figure 3 is a flowchart of another health retail supply chain optimization method according to an embodiment of the present invention;

[0056] Figure 4 is a flow chart of another health retail supply chain optimization method according to an embodiment of the present invention;

[0057] Figure 5 is a schematic diagram of a recommended process in a health retail supply chain optimization method according to an embodiment of the present invention;

[0058] Figure 6 is a structural block diagram of a health retail supply chain optimization device according to an embodiment of the present invention;

[0059] Figure 7 Schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of the present invention.

[0061] In existing technologies, some solutions use collaborative filtering algorithms to recommend products, but they only rely on user historical behavior data and are unable to tap into potential demand. Some solutions use natural language processing technology to analyze user reviews, but do not combine multimodal data (such as videos, text, orders, etc.) for comprehensive analysis. Existing marketing strategies mostly use a unified coupon issuance mechanism and lack dynamic analysis of user price sensitivity, making it difficult to achieve accurate recommendations and personalized marketing.

[0062] An embodiment of the present invention provides a health retail supply chain optimization method, which integrates multi-source heterogeneous data, uses a large model to explore users' potential needs, optimizes the product selection pool, and combines user portraits and price sensitivity analysis to achieve accurate recommendations and personalized marketing.

[0063] According to an embodiment of the present invention, an embodiment of a healthy retail supply chain optimization method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0064] In this embodiment, a health retail supply chain optimization method is provided, which can be used in the above-mentioned computer equipment, such as Figure 4 As shown, the computer device is provided with a data layer, a processing layer, an AI model layer and an application layer. Figure 1 FIG. 1 is a flow chart of a method for optimizing a health retail supply chain according to an embodiment of the present invention. Figure 1 As shown, the process includes the following steps:

[0065] Step S101: Acquire text data and user behavior data of a health retail scenario.

[0066] Specifically, the data layer acquires multi-source heterogeneous data from social platforms, including short video data (such as user viewing time and number of likes), historical order data (such as purchase and return records), external data (such as market trends and competitor product prices), and other user behavior data, as well as text data on social media topics. The social platforms include live broadcast rooms, short videos, and social media posts.

[0067] Text data is raw, unstructured text information, including but not limited to:

[0068] 1. User-generated content (UGC), including: short video comments / barrage (e.g., “This vitamin is really effective for improving sleep!”), social media posts or comments (e.g., “I’ve been suffering from insomnia lately. Please recommend a sleep aid product”), and product reviews (e.g., “Probiotics are effective, but the price is a bit high”).

[0069] 2. Operational texts, including health knowledge articles (such as the "Autumn Health Guide" released by the platform) and live broadcast scripts (such as the "Efficacy of Collagen Oral Liquid" introduced by the anchor), etc.

[0070] It should be noted that before using natural language processing technology to perform multi-level semantic analysis on text data, it is also necessary to use data cleaning tools (such as Python's Pandas library) to deduplicate, fill in missing values ​​and standardize the acquired short video data (such as user viewing time, number of likes), historical order data (such as purchase records, return records) and external data (such as market trends, competitor prices) and other user behavior data as well as the topic text data.

[0071] Step S102: Use natural language processing technology to perform multi-level semantic analysis on the text data to extract keywords of potential user needs.

[0072] Specifically, natural language processing technologies include NER (Named Entity Recognition), GPT-4 model (Generative Pre-trained Transformer 4, multimodal large language model) and LDA (Latent Dirichlet Allocation, topic model).

[0073] The processing layer uses NER, GPT-4, and LDA to perform multi-level semantic analysis on text data to extract keywords for users' potential needs. Keywords for users' potential needs include:

[0074] Explicit demand keywords: directly mentioned health needs (such as "insomnia" and "hair loss").

[0075] Implicit demand keywords: demand inferred from context (e.g., frequent discussion of “overtime” may imply demand for “liver protection products”).

[0076] Step S103: Build a user profile based on the user's potential demand keywords and user behavior data.

[0077] Specifically, user behavior data includes the short video data mentioned in step S101 above (such as user viewing time, number of likes), historical order data (such as purchase records, return records) and external data (such as market trends, competitor product prices).

[0078] The processing layer builds user portraits based on the extracted explicit demand keywords and implicit demand keywords, as well as short video data (such as user viewing time and number of likes), historical order data (such as purchase records and return records) and external data (such as market trends and competitor product prices).

[0079] User portraits include dimensions such as user age, gender, health needs, and spending power.

[0080] Step S104: Group users based on preset grouping criteria and user portraits in the health scenario, screen out corresponding health products for different user groups, and optimize and adjust the health products in demand based on market trend data in user behavior data to obtain an adjusted product selection pool.

[0081] Specifically, the AI ​​model layer groups users based on preset grouping criteria and user portraits in health scenarios, and uses clustering algorithms (such as the K-means algorithm) to screen suitable health products for different groups to form an initial product selection pool. The application layer then dynamically adjusts the initial product selection pool based on market trend data in user behavior data to ensure that the products meet market demand and obtain an adjusted product selection pool.

[0082] Step S105: Calculate the matching degree between the user and the preset candidate social platforms based on the user portrait and the product selection pool, and generate a social platform recommendation list based on the matching degree.

[0083] Specifically, the AI ​​model layer generates a multimodal large model based on user portraits and product selection pools, and the multimodal large model includes a recommendation model.

[0084] The application layer adopts a recommendation model and uses collaborative filtering algorithm to calculate the matching degree between users and live broadcast rooms and generate a recommended list of live broadcast rooms.

[0085] The health retail supply chain optimization method provided in this embodiment obtains text data and user behavior data of health retail scenarios, innovatively integrates multi-source data such as text data and user behavior data, and provides a comprehensive data foundation for the optimization of the health retail supply chain. Natural language processing technologies (such as NER, GPT-4 and LDA) are introduced to perform in-depth semantic analysis of text data, breaking through the limitations of traditional methods and accurately extracting keywords of users' potential needs. Multi-dimensional user portraits are constructed in combination with user behavior data to achieve a more comprehensive and in-depth understanding of user needs. Users are scientifically grouped, and suitable health products are screened according to the characteristics of different groups to improve the accuracy of product selection. The product selection pool is dynamically adjusted in combination with market trend data to make the product selection more in line with market dynamic needs. The matching degree between the user and the preset candidate social platform is calculated based on the user portrait and the product selection pool, and a personalized social platform recommendation list is generated based on the matching degree, which enables health products to accurately reach target users on appropriate social platforms, improve promotion efficiency, reduce marketing costs, and achieve accurate recommendation of social platforms, solving the problem that it is difficult to achieve accurate recommendation of health products and social platforms in the existing technology.

[0086] In this embodiment, a health retail supply chain optimization method is provided, which can be used in the above-mentioned computer device. Figure 2FIG. 1 is a flow chart of a method for optimizing a health retail supply chain according to an embodiment of the present invention. Figure 2 As shown, the process includes the following steps:

[0087] Step S201: Acquire text data and user behavior data of health retail scenarios. Figure 1 Step S101 of the illustrated embodiment will not be described in detail here.

[0088] Step S202: Use natural language processing technology to perform multi-level semantic analysis on the text data to extract keywords of potential user needs.

[0089] Specifically, the above step S202 includes:

[0090] In step S2021, natural language processing technology is used to perform sequence labeling, sentiment polarity analysis, and topic distribution extraction on the text data, respectively, to identify user intentions, health topic attitudes, and topic distribution.

[0091] Specifically, the natural language processing technology in this embodiment includes NER (Named Entity Recognition), GPT-4 model (Generative Pre-trained Transformer 4, multimodal large language model) and LDA (Latent Dirichlet Allocation, topic model).

[0092] 1. Intent Recognition (Sequence Labeling):

[0093] Technical implementation: Use NER technology to extract health entities in text.

[0094] Example: Input the text "I've been staying up late lately and my skin has gotten worse" → Output entity tags, output health entity tags "staying up late" and "skin".

[0095] 2. Sentiment Polarity Analysis

[0096] Technical implementation: Use GPT-4's fine-grained sentiment analysis module to output values ​​in the range [-1,1].

[0097] Example: If the user text data is "This probiotic is generally effective and not cost-effective," the GPT-4 model identifies -0.3 (negative sentiment).

[0098] 3. Topic modeling (LDA).

[0099] Technical implementation: Topic distribution is extracted based on the Latent Dirichlet Allocation (LDA) model.

[0100] Example: User text data is ["Insomnia Melatonin Recommendation", "Probiotics for Gastrointestinal Conditioning", "Collagen for Skin Elasticity"]. The LDA model is used to identify Topic 1: Insomnia, Melatonin, Anxiety; Topic 2: Probiotics, Gastrointestinal, Digestion, etc.

[0101] Step S2022: Generate explicit demand keywords and implicit demand keywords based on user intention, attitude towards health topics, topic distribution, and preset health knowledge graph.

[0102] Specifically, explicit demand keywords and implicit demand keywords are generated based on user intention, attitude towards health topics, topic distribution, and preset health knowledge graph.

[0103] For example:

[0104] Original text data: "I have a lot of work pressure and drink coffee every day to stay awake, but I have started to lose my hair recently."

[0105] Extraction process:

[0106] a. Named Entity Recognition: Identify health entities such as "work stress," "coffee," and "hair loss."

[0107] b. Associated with the preset health knowledge graph:

[0108] "Work pressure" → related to keywords such as "anxiety" and "sleep quality" in the health knowledge map.

[0109] "Coffee" → Related keywords such as "Vitamin B consumption" in the health knowledge map.

[0110] "Hair loss" → related to keywords such as "hair loss" and "hair care nutrients" in the health knowledge map.

[0111] c. Output the final keywords: ["anxiety", "B vitamins", "hair loss"]. "Anxiety" and "hair loss" are explicit demand keywords, while "B vitamins" is an implicit demand keyword.

[0112] Step S203: Build a user profile based on the user's potential demand keywords and user behavior data.

[0113] Specifically, user behavior data includes social platform short video data, historical order data, social platform interaction data, and external data; the types of user behavior data are shown in Table 1 below:

[0114] Table 1 User behavior data types

[0115]

[0116] The above step S203 includes:

[0117] Step S2031: Obtain user basic attribute characteristics and extract health demand characteristics based on user potential demand keywords.

[0118] Specifically, we use the Z-Score normalization method to standardize the numerical features in user behavior data, and use one-hot encoding or embedding to represent categorical features. In Table 1 above, video classification and purchase category are categorical features, while the rest are continuous features.

[0119] For example, if there are three categories of category features, such as fruits and vegetables, snacks and beverages, and health products, then the one-hot encoding is divided into three columns. If fruits and vegetables have been purchased, the fruit and vegetable field will be marked as 1;

[0120] Continuous features such as video viewing time and average order value will be normalized as a whole to conform to the normal distribution.

[0121] User basic attribute characteristics include user age, gender, location, etc.

[0122] Extract health demand features based on user potential demand keywords, such as sleep quality, gastrointestinal health, skin condition, etc.

[0123] For example:

[0124] Basic attribute characteristics: {Age: 25-30 years old, Gender: Female, Region: Beijing}.

[0125] Health needs characteristics: {sleep quality: poor, gastrointestinal health: sensitive, skin condition: dry}.

[0126] Step S2032: extracting consumer demand characteristics based on social platform short video data, historical order data, social platform interaction data, and external data.

[0127] Consumer demand characteristics include price sensitivity, category preference, and purchase frequency.

[0128] For example, the consumer demand characteristics are: {price sensitivity: high, category preference: health products, purchase frequency: once a month}.

[0129] The basic attribute characteristics, health demand characteristics and consumption demand characteristics are used as the portrait label system, namely:

[0130] Basic attribute characteristics: {Age: 25-30 years old, Gender: Female, Region: Beijing}.

[0131] Health needs characteristics: {sleep quality: poor, gastrointestinal health: sensitive, skin condition: dry}.

[0132] Consumer demand characteristics: {price sensitivity: high, category preference: health products, purchase frequency: once a month}.

[0133] Step S2033: Build a user profile based on basic attribute characteristics, health demand characteristics, and consumption demand characteristics.

[0134] Specifically, a layered fusion strategy is adopted to integrate three types of features into a dynamic portrait system, including:

[0135] Feature vector generation: Basic attributes, health needs, consumer demand characteristics, and unstructured text features (such as health reviews) are used to generate low-dimensional vectors using context-aware embedding technology (BERT can be used in this embodiment); for structured behavioral data (such as order amount), they are normalized and mapped to latent space vectors.

[0136] Portrait fusion: Dynamically weight basic attributes, health needs, and consumer demand vectors through multimodal feature fusion models (such as neural networks and graph attention networks), where the weight coefficients are calculated in real time based on the user's most recent N behavioral data.

[0137] Portrait update: Establish a time sliding window mechanism, set the frequency of monthly updates of basic attributes, weekly updates of health needs, and real-time updates of consumer needs to ensure the timeliness of the portrait.

[0138] Visual presentation: Use tools such as Tableau to convert the portrait into a multi-dimensional label system (such as "25-30 years old / female / fitness enthusiast / high spending power / price sensitive").

[0139] It should be noted that the relationship between keyword extraction and user portraits includes:

[0140] (1) Logical association:

[0141] Keyword extraction: Mining users’ explicit / implicit health needs (such as “insomnia” and “hair loss”) from unstructured text.

[0142] User portrait construction: Map keywords to a structured tag system and integrate them with behavioral data to form a comprehensive portrait.

[0143] (2) Examples of synergy:

[0144] Scenario: User A mentioned in a comment that "building muscle through fitness is difficult" and watched protein powder reviews several times in a short video.

[0145] Processing flow: a. Text analysis extracts keywords: ["fitness", "muscle gain", "protein"]; b. Behavioral data analysis: 90% completion rate for protein powder-related videos, indicating purchase of whey protein; c. Image synthesis results: {"fitness and muscle gain": "high demand", "sports nutrition product preference": "whey protein", "price sensitivity": "low"}.

[0146] Application: Recommend high-protein healthy foods and issue “50 off for purchases over 500” coupons (for price-sensitive users).

[0147] (3) The comparison of technical effects is shown in Table 2 below:

[0148] Table 2 Comparison of technical effects

[0149] method Only behavioral data Combining keywords and behavioral data Demand identification accuracy Identify explicit needs that have been purchased (60%) Capturing hidden needs that haven’t been purchased (85%) Recommendation conversion rate 4.2％ 7.8％ Long-tail product coverage 15％ 38％

[0150] Step S204: Group users based on the preset grouping criteria and user portraits in the health scenario, filter out the corresponding health products for different user groups, and optimize and adjust the health products in demand based on the market trend data in the user behavior data to obtain the adjusted product pool. Figure 1 Step S104 of the embodiment shown is not described here in detail. Step S205: Calculate the matching degree between the user and the preset candidate social platforms based on the user portrait and the product selection pool, and generate a social platform recommendation list based on the matching degree. Figure 1 Step S105 of the illustrated embodiment will not be described in detail here.

[0151] The health retail supply chain optimization method provided in this embodiment can deeply mine users' potential needs by acquiring text data and user behavior data of health retail scenarios and performing multi-level semantic analysis using natural language processing technology. The implicit needs mined from text data complement each other with behavioral data, focusing on both the user's surface behavior and in-depth insights into internal needs, ensuring that the user portrait accurately covers all aspects of the user's characteristics. The basic attribute characteristics, health demand characteristics, and consumer demand characteristics are extracted separately to analyze users from different dimensions. The basic attribute characteristics clarify the user's basic background, the health demand characteristics focus on the user's health demands, and the consumer demand characteristics show the user's consumption behavior pattern. The fusion of multi-dimensional features makes the user portrait three-dimensional and full, and can deeply portray the user's image.

[0152] In this embodiment, a health retail supply chain optimization method is provided, which can be used in the above-mentioned computer device. Figure 3 FIG. 1 is a flow chart of a method for optimizing a health retail supply chain according to an embodiment of the present invention. Figure 3 As shown, the process includes the following steps:

[0153] Step S301: Acquire text data and user behavior data of health retail scenarios. Figure 2 Step S201 of the illustrated embodiment will not be described in detail here.

[0154] Step S302: Use natural language processing technology to perform multi-level semantic analysis on the text data to extract keywords of potential user needs. Figure 2 Step S202 of the illustrated embodiment will not be described in detail here.

[0155] Step S303: Build a user profile based on the user's potential demand keywords and user behavior data. Figure 2 Step S203 of the illustrated embodiment will not be described in detail here.

[0156] Step S304: Group users based on preset grouping criteria and user portraits in the health scenario, screen out corresponding health products for different user groups, and optimize and adjust the health products in demand based on market trend data in user behavior data to obtain an adjusted product selection pool.

[0157] Specifically, the preset clustering standards include health demand characteristic standards, consumption behavior characteristic standards, health status characteristic standards and time and space characteristic standards; Figure 4 As shown, the above step S304 includes:

[0158] Step S3041, based on health demand characteristic standards, consumption behavior characteristic standards, health status characteristic standards, spatiotemporal characteristic standards and user portraits, a clustering algorithm is used to group users to obtain different user groups.

[0159] Specifically, in health retail scenarios, user segmentation criteria must be combined with multi-dimensional health characteristics, consumer behavior patterns, and market dynamics to achieve precise segmentation through the following technical solutions, including:

[0160] 1. Clustering criteria and characteristic system are shown in Table 3 below:

[0161] Table 3 Clustering criteria and characteristic system

[0162]

[0163]

[0164] 1. Health Needs Characteristics: Based on the health needs characteristics of user profiles, we further refine specific needs such as disease prevention, rehabilitation and health care, and nutritional supplements. For example, we can categorize groups with different disease prevention needs based on their interest in topics related to diseases like hypertension and diabetes, as well as their inquiries about products like antihypertensive medications and blood glucose meters. We can also identify groups with rehabilitation and health care needs based on their browsing preferences for content like postoperative rehabilitation and chronic disease management.

[0165] 2. Consumer Behavior Characteristics: Based on consumer demand characteristics, we set segmentation indicators such as consumption frequency (high, medium, low), average order value (high, medium, low), and brand loyalty (high, medium, low). For example, users who frequently purchase with high average order value and consistently choose a specific brand are classified as high-value and brand-loyal consumers; users who purchase infrequently and tend to try different brands are classified as price-sensitive and brand-exploring.

[0166] 3. Health Status Characteristic Standards: Combining basic user attributes and health needs data, users' health status is categorized as healthy, sub-healthy, and sick. Health status is determined by the user's interest in physical examination reports, health monitoring device data, and preference for health education content. For example, users who focus on healthy eating, exercise, and fitness and have no disease-related search history are considered healthy; users who frequently search for symptoms and relief methods for fatigue, insomnia, and other symptoms are classified as sub-healthy.

[0167] 4. Spatial and temporal characteristics: In the temporal dimension, factors such as seasons, holidays, and peak consumption periods are taken into consideration. For example, demand for warmth and nourishing health products increases in winter, and sales of gift-type health products increase during holidays. In the spatial dimension, the user's region is divided based on differences in climate, economic level, and medical resources. For example, northern regions have a higher demand for heating equipment and frostbite prevention and treatment products, while economically developed regions have a stronger demand for high-end health management services.

[0168] Select an appropriate clustering algorithm based on data characteristics and clustering requirements. The K-means algorithm is suitable for situations where data distribution is relatively uniform and clusters are approximately spherical, and can quickly and efficiently divide users into a specified number of groups. The hierarchical clustering algorithm is suitable when the number of clusters is uncertain. It displays the hierarchical structure of the data by constructing a clustering tree, gradually determining the optimal number of clusters. The DBSCAN algorithm is robust to noisy data and is suitable for scenarios where data distribution is irregular and outliers are present, accurately identifying core user groups and marginal users.

[0169] Taking the K-means algorithm as an example, the optimal number of clusters K is determined by combining methods such as the Elbow Method and Silhouette Coefficient with business objectives (such as covering at least five major categories of health needs). The Elbow Method plots the clustering error (such as the sum of squares within the cluster) curve under different K values ​​and finds the inflection point of the curve as the optimal K value; the Silhouette Coefficient calculates the silhouette coefficient of each sample point, and the K value corresponding to the maximum overall silhouette coefficient is the optimal number of clusters to ensure the rationality and effectiveness of the clustering results. The processed data and selected features are input into the clustering algorithm to perform the clustering operation. The algorithm calculates the cosine similarity or difference between users based on the set standards and distance metrics (such as Euclidean distance and Manhattan distance), divides similar users into the same group, and ultimately obtains different user groups.

[0170] For non-convex distribution data, DBSCAN clustering algorithm or spectral clustering algorithm can also be used. The user grouping results are shown in Table 4 below:

[0171] Table 4 User grouping results

[0172]

[0173] Step S3042: extracting trend feature vectors from the market trend data, and inputting the trend feature vectors into a preset time series prediction model for training to obtain trend prediction results.

[0174] Specifically, 1. Market trend data types and collection methods are shown in Table 5 below:

[0175] Table 5 Market trend data types and collection methods

[0176]

[0177] Based on market trend data, key quantitative features are extracted, such as the growth rate of the health product market size, the month-on-month changes in monthly sales of subdivided categories (such as probiotics and vitamins), and the price fluctuations of competing products.

[0178] For text data such as policy documents, news information, etc., natural language processing technology (NLP) is used to convert the text into vectors through word embedding (such as BERT), and the word frequency of keywords (such as "health consumption upgrade" and "traditional Chinese medicine policy support") is extracted as features.

[0179] Combine quantitative indicators and text features to form a multi-dimensional trend feature vector. For example, the market growth rate of a certain type of health product (numerical) and the policy keyword vector (text) are combined into a complete feature vector to fully reflect market trends.

[0180] Example: The trend feature vector of a vitamin C product includes: average daily search volume, number of social media mentions, price reduction of competing products (compared to the current social platform), CDC influenza alert level (level 1-5), and policy risk probability (model prediction value).

[0181] Select an appropriate time series forecasting model based on the data characteristics. For example, the traditional ARIMA model (Autoregressive Integrated Moving Average Model) is suitable for linear trend data, while deep learning models such as LSTM (Long Short-Term Memory Network) and Prophet are better at fitting nonlinear and complex trends. If the data exhibits seasonal or cyclical variations, the Holt-Winters seasonality model can be used.

[0182] Taking the Prophet model as an example, the historical trend feature vectors and the corresponding actual data are divided into training and test sets in a 7:3 ratio. The training set is used to learn model parameters, and the test set is used to evaluate the model's generalization ability. Mean squared error (MSE) is used as the loss function to measure the difference between the predicted and actual values. The Adam optimizer is used to adjust model parameters, and the weights are updated through the backpropagation algorithm to reduce the loss function value. Model performance is evaluated using metrics such as mean absolute error (MAE) and root mean square error (RMSE). By adjusting model hyperparameters (such as the number of LSTM layers and learning rate) or trying different model architectures, forecast accuracy is optimized to obtain trend forecast results.

[0183] Example: Take sales prediction model training as an example. Features include date, sales volume, trend, etc. External regression factors are added during the prediction process to generate sales forecast results for the next 14 days.

[0184] Step S3043: Adjust the selection weights of in-demand health products based on trend prediction results and preset product selection decision rules.

[0185] Specifically, the preset product selection decision rules are formulated, that is, multi-dimensional decision rules are formulated, including:

[0186] Market potential rule: If the predicted growth rate of the market size of a certain type of product exceeds a threshold (such as 15%) and the company has sufficient existing inventory, the selection weight of this product will be increased.

[0187] Competitive strategy rules: When a competitor launches a similar product at a lower price than your product, reduce the weight of that product and increase the weight of the differentiated product.

[0188] User demand rules: Combined with the user portrait segmentation results, if the demand for a certain type of product increases among high-value user groups, increase its product selection weight.

[0189] The trend prediction results are fed into the product selection decision rule engine, and the rules are matched one by one. For example, if a certain type of health food is predicted to see significant sales growth in the future and meets the market potential rule, the product selection weight for that product will be increased from its initial value (e.g., 1.0) to 1.3.

[0190] For products that do not comply with the rules, their weight will be reduced accordingly. For example, if market demand for a product decreases and inventory is high, its weight will be reduced from 1.0 to 0.7, reducing subsequent purchases.

[0191] For example, the formula for calculating product selection weight is: weighted sum (0.6*basic score (trend_score) + 0.4*market trend score) / inventory factor.

[0192] Step S3044: Dynamically update and remove in-demand health products based on the adjusted weights to obtain an adjusted product selection pool.

[0193] Specifically, based on trend forecast results and weight adjustments, new products with increased weights that align with corporate strategy (e.g., predicted increased demand for smart health monitoring devices) are added to the product selection pool after undergoing processes like supplier evaluation and sample testing. For existing products, if their weights increase, inventory levels are increased and product page presentation is optimized (e.g., highlighting product advantages and increasing user reviews). If their weights decrease, inventory is reduced or cleared through promotional activities.

[0194] Delisting thresholds are set. For example, if a product's selection weight falls below 0.5 and sales fall below expectations for two consecutive months, delisting is triggered. Risk assessments are conducted before delisting, taking into account factors like inventory overstock costs and customer satisfaction. If inventory is high, priority is given to clearing out inventory through limited-time discounts or bundle sales before delisting, ultimately forming a dynamically updated selection pool.

[0195] The actual application examples of steps S3041 to S3044 are as follows:

[0196] Scenario: In October 2023, the system detected the following trend signals:

[0197] 1. The 7-day growth rate of topics related to "Mycoplasma pneumonia" on social media was 320%.

[0198] 2. The search volume for "lactoferrin" on competing platforms increased by 200%, and competing products began to reduce prices and promote sales.

[0199] 3. Meteorological data show that PM2.5 concentrations in many parts of the north have exceeded 150.

[0200] 4. Trend Analysis: Predicting a surge in demand for "lung-nourishing" and "immunity-boosting" health products, calculate the weight increase coefficient for related categories:

[0201] 1) Lactoferrin: trend_score=0.92→weight+35%.

[0202] 2) Air purifier: trend_score = 0.85 → weight + 28%.

[0203] 5. Product selection pool adjustment:

[0204] Newly added: Chuanbei loquat paste (a traditional Chinese medicine for moistening the lungs) and HEPA filter air purifier.

[0205] Improvement: Lactoferrin capsules rose from 15th to 3rd place; vitamin C effervescent tablets rose from 8th to 5th place.

[0206] Demotion: The weight of summer sunscreen products is reduced by 60%.

[0207] 6. Supply chain collaboration: Send urgent purchase orders to suppliers to secure lactoferrin raw materials; adjust logistics priorities to ensure that the northern warehouse is replenished within 3 days.

[0208] 7. Marketing linkage: The live broadcast room pushes the "enhancing immunity" topic, and the anchor's speech embeds trend keywords; "air purifier + mask" combination coupons are issued to users in high-risk areas.

[0209] Step S305: Calculate the matching degree between the user and the preset candidate social platforms based on the user portrait and the product selection pool, and generate a social platform recommendation list based on the matching degree.

[0210] Specifically, the recommended process diagram in the health retail supply chain optimization method is as follows: Figure 5 As shown, the above step S305 includes:

[0211] Step S3051: Use embedding technology to convert the user profile into a dense vector.

[0212] Collect and integrate basic user attributes (age, gender, region, etc.), health needs characteristics (potential demand keywords extracted through text semantic analysis), and consumption needs characteristics (extracted from short videos on social platforms, historical orders, and other data). Use Python's Pandas library for data cleaning, employing multiple imputation methods for missing values, such as filling in age and consumption amount based on the mean or median of the user's group. Use regular expressions to unify data formats, such as standardizing dates and region codes.

[0213] The BERT model, based on the Transformer architecture, feeds textual features from user profiles (such as health need keywords and review content) into the model, extracting contextual semantic vector representations at the character / word level. These are then aggregated into fixed-length vectors through pooling operations (such as average pooling and max pooling). Numerical features (such as consumption frequency and average order value) are normalized to the range [0, 1] before being concatenated with the text vector.

[0214] Graph embedding technology: If user profiles have complex relationships (such as the purchasing relationship between users and health products, or the interactive relationship between users on social platforms), a heterogeneous graph network of users, products, and social behaviors is constructed. Graph embedding algorithms such as GraphSAGE and GAT are used to learn the low-dimensional representation of nodes in the graph structure, ultimately obtaining a dense vector containing multi-dimensional information about users.

[0215] Step S3052: extract corresponding platform features of the preset candidate social platforms and represent the platform features using vectorization; the platform features include platform content, platform products, and real-time performance.

[0216] Specifically, text content processing: For health-related articles, user comments, and other text content on social platforms, tools such as NLTK (Natural Language Toolkit) and spaCy (a high-performance NLP (Natural Language Processing) library) are used for word segmentation, part-of-speech tagging, and named entity recognition to extract health keywords (such as "weight loss recipes" and "probiotic benefits"). Keyword weights are calculated using the TF (Term Frequency)-IDF (Inverse Document Frequency) algorithm, and then keywords are converted into word vectors using Word2Vec (a technology that converts words into numerical vectors (word embeddings)) or FastText (a text classification and word vector generation tool). Finally, the feature vector of the text content is obtained through weighted summation or average pooling.

[0217] Multimedia content processing: Computer vision technology is used for multimedia content such as short videos and images. For example, a pre-trained ResNet (Residual Network) model is used to extract image features from video keyframes, and a VGGish model (audio classification model) is used to extract audio features. These features are then mapped to a fixed dimension through a fully connected layer and concatenated with the text content feature vector to form a complete platform content feature vector.

[0218] For example, taking the live broadcast room on a social platform as an example, the live broadcast room features include the live broadcast theme extracted by NLP, the associated healthy products in the selection pool, the average viewing time (seconds), the historical conversion rate, and the anchor's professionalism rating (user evaluation).

[0219] The embedding technology is used to vectorize the above live broadcast room features to obtain feature vectors.

[0220] Step S3053: A hybrid collaborative filtering algorithm is used based on the product selection pool to calculate multiple matching degrees between the dense vector and the corresponding platform feature vector.

[0221] Specifically, we combine user-based collaborative filtering (CF) with item-based collaborative filtering (CF), and incorporate content-based filtering. We first use cosine similarity to calculate content-based matching between dense user profile vectors and social platform feature vectors, measuring their similarity at the feature level. We then use historical user behavior data on various social platforms (such as browsing time and interaction frequency) to calculate user-based collaborative filtering matching. Furthermore, we calculate item-based collaborative filtering matching based on the sales performance of products in the selection pool across different social platforms.

[0222] Through experimentation or experience, we set weights for the matching degrees of different algorithms. For example, we set a weight of 0.4 for content-based matching, 0.3 for user-based collaborative filtering, and 0.3 for item-based collaborative filtering. We then take the weighted sum of these three matching degrees to determine the final matching degree between each user and the preset candidate social platforms.

[0223] Step S3054: Arrange the multiple matching degrees in descending order, and generate a social platform recommendation list based on the matching degrees after the descending order and the social platform operation strategy.

[0224] Specifically, the calculated multiple matching degrees are arranged in descending order to form an ordered matching degree list.

[0225] If a social platform is running a special promotion for health products, we'll give it a certain weight bonus to increase its priority in the recommendation list. For high-value user groups (such as those with high spending power and high activity), we'll prioritize social platforms with strong professionalism and high user stickiness. For new or low-activity users, we'll recommend social platforms with high traffic and rich content to improve user experience and engagement.

[0226] Based on the adjusted matching order, several top-ranked social platforms are selected to generate a personalized social platform recommendation list, which is provided to users or used for precision marketing by enterprises. Step S3055: Real-time user behavior on the social platform is obtained, and the social platform recommendation list is dynamically adjusted based on the real-time user behavior.

[0227] Specifically, real-time behavioral data collection includes:

[0228] Interactive behavior: Capture user actions such as likes, comments, reposts, and favorites on social platforms, and record the operation objects (such as health posts and product links), operation timestamps, and operation frequency.

[0229] Browsing behavior: Monitor the types of pages users browse (product details pages, health science articles, user dynamics), duration of stay, and scrolling depth, and obtain data in real time through page embedding technology.

[0230] Search behavior: Obtain the search keywords entered by users (such as "weight loss methods" and "vitamin recommendations"), the number of searches, and clicks on search results.

[0231] Transaction behavior: If jump purchase is involved, track actions such as adding to the shopping cart, placing an order, and making a payment, and record the transaction product, transaction amount, and transaction time.

[0232] Assign a higher weight to real-time user behavior data. For example, assign a weight of 0.8 to behavior data within the last hour, while reducing the weight of historical behavior data to 0.2. When calculating the matching degree between the user profile and the social platform feature vector, the real-time behavior feature vector is fused with the original user profile vector according to the weights to obtain the updated user representation vector.

[0233] Differentiated weights are set for different types of real-time behaviors. For example, purchasing behavior is weighted higher than browsing behavior, and the weight of comment behavior is adjusted according to emotional tendencies (positive comments have higher weights, and negative comments may reduce the weight of related recommendations).

[0234] Using online learning algorithms (such as FTRL-Proximal), when new real-time behavior data arrives, only the affected model parameters are locally updated, avoiding the time overhead of retraining the entire model. For example, when a user generates a new search behavior, the model parameters related to the search keyword are updated to optimize search result recommendations. A memory network is used to store recent user behavior sequences. When making model predictions, historical memory is combined with current real-time behavior to better capture the dynamic changes in user behavior.

[0235] The updated user representation vector and the social platform feature vector are then recalculated using a hybrid collaborative filtering algorithm, specifically considering the impact of real-time behavior on the matching degree. For example, if a user's real-time behavior indicates a strong interest in a particular type of health livestream, the matching degree of the social platforms containing that type of livestream will be increased. The original recommendation list is then reordered in real time based on the new matching degree, prioritizing the display of highly compatible social platforms. Furthermore, filtering rules are set up so that if a social platform has recently experienced a surge in negative reviews or experienced service outages, it will be temporarily removed from the recommendation list.

[0236] Using a waterfall loading method, the backend continuously monitors user behavior as they browse the recommendation list. When it detects that a user lingers for an extended period of time or frequently returns to a particular recommendation, it dynamically inserts relevant supplementary recommendations (such as similar social platforms and trending health topics). User actions on the adjusted recommendation list (such as clicking on new recommendations or ignoring certain recommendations) are recorded and used as a basis for the next round of recommendation adjustments.

[0237] Step S306: Obtain the user's historical consumption behavior and price sensitivity; in the process of generating the recommendation list on the social platform, based on the historical consumption behavior, price sensitivity and real-time user behavior, use operations optimization theory to formulate a personalized coupon issuance strategy; use A / B testing algorithm to optimize the personalized coupon issuance strategy.

[0238] Specifically, the implementation of a personalized coupon strategy is a dynamic decision-making optimization problem that requires balancing short-term conversions with long-term user value. This embodiment uses Deep Reinforcement Learning (DRL) combined with constrained optimization. The specific process is as follows:

[0239] Data collection → state representation → policy network → action (coupon issuance) → environment feedback → policy update.

[0240] 1. Data collection includes:

[0241] Historical consumption behavior includes: purchase frequency, coupon usage rate, and click-through conversion rate.

[0242] Price sensitivity: calculated using a logistic regression model (0-1 range).

[0243] User real-time behavior: product browsing and duration of stay in the live broadcast room in the last 30 minutes.

[0244] 2. State representation: Convert historical consumption behavior, price sensitivity, and real-time user behavior into a state vector, where historical consumption behavior is statistically represented as a 10-dimensional vector, price sensitivity is a 1-dimensional scalar, and real-time user behavior is embedded as a 32-dimensional vector.

[0245] 3. Strategic Network includes:

[0246] 1. Coupon action definition: denomination (5 yuan, 10 yuan, 20 yuan, 30 yuan), validity period (24 hours, 48 ​​hours, 72 hours). Action encoding: One-hot encoding is combined into 54 possible actions (3 types × 4 denominations × 3 validity periods).

[0247] 2. Multi-objective reward function:

[0248] Instant rewards: include conversion rewards (users complete purchases) and revenue rewards (based on order amount).

[0249] Long-term value penalty: Prevent over-reliance on high-discount strategies and maintain long-term user value.

[0250] Sensitivity adaptive rewards: Provide appropriate discounts to price-sensitive users to increase conversion rates.

[0251] 3. Policy network architecture:

[0252] Actor-Critic framework: learns the policy function (Actor) and the value function (Critic) simultaneously.

[0253] Input layer: 43-dimensional state features (user profile, historical behavior, etc.).

[0254] Output layer: 54-dimensional action space (different coupon combination strategies).

[0255] 4. Constrained Optimization Integration:

[0256] Budget constraint: Ensure that the total daily coupon cost does not exceed the preset budget.

[0257] Lagrange multiplier method: Incorporate constraints into the reinforcement learning loss function to achieve constrained optimization.

[0258] This design balances short-term benefits and long-term value through multi-objective rewards, and combines constraint optimization to ensure effective resource utilization. It is particularly suitable for optimizing personalized coupon distribution strategies in health retail scenarios and realizing coupon distribution.

[0259] 4. Technical implementation details of A / B testing to optimize marketing strategies:

[0260] 1) Overall design of A / B testing:

[0261] 1. Test objectives, including:

[0262] Core objective: To verify the effectiveness of new marketing strategies (such as reinforcement learning coupon strategies) in improving conversion rate (CVR), average order value (AOV), user retention rate, and other indicators compared to baseline strategies (such as manual rule strategies).

[0263] Auxiliary goal: Identify the differentiated impact of the strategy on different user groups (such as price-sensitive users vs. high-net-worth users).

[0264] 2. Experimental group design, using stratified randomization to ensure comparability between groups:

[0265] a. Stratified randomization: supports stratification based on multiple factors (price sensitivity, consumption level, health needs); ensures that users within each stratum are evenly distributed into the test and control groups.

[0266] b. Deterministic allocation: Users with the same user_id and strata combination are always assigned to the same group. This eliminates the need to store historical allocation results, ensuring reproducible allocation results.

[0267] c. Uniform distribution: Achieve an approximate 50%:50% distribution ratio through hash value modulus operation; the randomness of the hash function ensures the balance of the grouping.

[0268] This grouping method is particularly suitable for personalized strategy testing in health retail scenarios. It can effectively control test variables and evaluate the effectiveness of different coupon strategies.

[0269] 3. The core indicators are defined as shown in Table 6 below:

[0270] Table 6 Core indicators

[0271] index Calculation formula Target improvement value Conversion Rate (CVR) Number of converted users / number of exposed users × 100% +15％ Average Order Value (AOV) Total transaction volume / number of completed orders +10％ User retention rate (7 days) Number of active users on the 7th day / number of exposed users on the first day × 100% +5％ ROI (Incremental revenue - preferential cost) / preferential cost × 100% >120％

[0272] 2. Dynamic Traffic Allocation and Multi-Armed Bandit Machine:

[0273] 1. Adaptive traffic distribution: Thompson sampling is used to dynamically adjust the grouping ratio, including:

[0274] a. Bayesian update mechanism:

[0275] Beta distribution is used as the prior and posterior for the success rate (conjugate prior property).

[0276] The initial parameters α=1, β=1 correspond to uniform distribution (no information prior).

[0277] After each observation of conversion / non-conversion, the α or β parameters are updated accordingly.

[0278] b. Exploration and Exploitation Balance:

[0279] Balance exploration and exploitation by sampling from a Beta distribution.

[0280] Early stage: Due to high uncertainty, sampling results fluctuate greatly, prompting the trial of different strategies (exploration).

[0281] Late stage: The posterior distribution of the well-performing strategies converges, and the probability of being selected increases (exploitation).

[0282] c. Online learning features:

[0283] Model parameters are updated immediately after each user decision.

[0284] There is no need to store all historical data, only the α and β parameters need to be maintained.

[0285] Naturally adapt to non-stationary environments (such as user behavior changing over time).

[0286] This approach is particularly suitable for coupon optimization in health retail scenarios. It can maximize short-term profits while testing different strategies and automatically converge to the optimal strategy as data accumulates.

[0287] 2. Multivariate testing, combining different coupon parameters:

[0288] Taking the typical orthogonal experimental design process as an example, the main features include:

[0289] a. Factor and level design:

[0290] Discount type (2 levels): Full-spend discount / Percentage discount. Discount value (3 levels): 5 yuan / 10 yuan / 20 yuan. Validity period (2 levels): 24 hours / 48 hours.

[0291] b. Orthogonal array generation:

[0292] Use the generate_orthogonal_array function to automatically generate the experimental design.

[0293] Using the L12 (2^1×3^1×4^1) mixed level orthogonal array, only 12 experiments are required.

[0294] Compared with a full-scale experiment (2×3×2=12 combinations), an orthogonal design can effectively reduce the number of experiments.

[0295] c. Advantages of experimental design:

[0296] Balanced distribution: Each level of each factor is evenly matched with the levels of other factors.

[0297] Independence of main effects: The influence of each factor on the outcome can be assessed independently.

[0298] Efficiency: Obtain key information through fewer experiments.

[0299] This experimental design method is particularly suitable for coupon strategy optimization in health retail scenarios. It can quickly find the key factor combination that affects user conversion while reducing testing costs.

[0300] 3. Statistical analysis methods:

[0301] 1. Hypothesis Testing:

[0302] Null hypothesis H: There is no significant difference between the new strategy and the old strategy.

[0303] Test method: T test (applicable to normal distribution indicators such as AOV).

[0304]

[0305] Among them, t is the T test value, and the p value can be obtained according to the t value to indicate whether there is a significant difference. and is the mean of the two groups being compared, s p is the combined standard error of the two groups, n A and n B is the sample size of each group.

[0306] Chi-square test (applicable to proportional indicators such as conversion rate):

[0307]

[0308] Among them, χ represents the chi-square value, which is used to measure the degree of difference between the actual value and the theoretical value (that is, the core idea of ​​the chi-square test). i represents the actual observed frequency, E i Indicates the expected frequency.

[0309] 2. Bayesian analysis: Calculate the Bayes factor (BF) to determine the pros and cons of the strategy:

[0310]

[0311] Among them, BF represents the Bayes factor, P(Data|H1) represents the probability of the current data appearing under the new model H1, and P(Data|H0) represents the probability of the current data appearing under the old model H0.

[0312] When BF>3, the new strategy is considered to be significantly better than the old strategy.

[0313] 3. Long-term effect evaluation, using the Autoregressive Integrated Moving Average (ARIMA) model to predict long-term impact:

[0314]

[0315] Among them, L represents the lag operator, which is used to analyze the time series impact of the strategy on user retention and is defined as L i y t =y t-i, indicating that the time series data moves back i time steps, Ly t =y t-1 (represents the observation value of the previous day); p represents the autoregressive coefficient, which indicates the order of the autoregressive (AR) term used in the model, that is, the value of the past p time points. For example, if p = 2, the model will consider y t-1 、y t-2 y t ; q represents the moving average order, which indicates the order of the moving average (MA) term in the model, that is, the error term (residual) that depends on the past q time points; for example, if q = 1, the model will consider the error ε of the previous time point t-1 ; i represents the index variable used to traverse the lag term of autoregressive or moving average (from 1 to p or from 1 to q); φ i is the autoregressive coefficient, which represents the weight of the i-th autoregressive term and measures the effect of y t-i For the current value y t The influence of θ i is the moving average coefficient, which represents the weight of the i-th moving average term and measures the historical error ε t-i For the current value y t The influence of ε t is the error term, representing the white noise at time point t (random error with mean 0 and constant variance); t is the value of the time series at time t.

[0316] 4. Optimize the feedback loop:

[0317] 1. Real-time effect monitoring dashboard, which uses a streaming computing framework (such as Apache Flink or Spark Streaming) to implement real-time A / B testing indicator calculation. Its main functions include:

[0318] a. Data Source and Processing:

[0319] Consumes real-time conversion events from the Kafka topic conversion_events.

[0320] Events include experimental groups (A / B groups) and user behavior data (whether conversion occurred, order amount).

[0321] b. Time window setting:

[0322] Data aggregation is performed using a 5-minute tumbling window.

[0323] Metrics are calculated independently for each window, without overlap.

[0324] c. Calculation of core indicators:

[0325] CVR (conversion rate): number of converted users / total number of exposed users.

[0326] AOV (Average Order Value): Total order amount / number of orders.

[0327] d. Real-time and high efficiency:

[0328] Data is processed event by event with low latency (response within seconds).

[0329] Incremental aggregation calculation with low resource consumption.

[0330] This real-time indicator calculation method is particularly suitable for coupon strategy optimization in health retail scenarios. It can provide timely feedback on the effectiveness differences of different strategies and support rapid decision-making adjustments.

[0331] 2. Dynamically tune policy parameters and use gradient descent optimization based on A / B testing results:

[0332]

[0333] Among them, θ t+1 represents the updated policy parameters for the next iteration or policy adjustment, θ t represents the strategy parameter vector at the tth iteration (e.g., recommendation algorithm weights, ad bid coefficients, etc.). In A / B testing, this may correspond to parameters of different strategy versions (e.g., UI design weights, recommendation model thresholds). η represents the learning rate, which controls the compensation for parameter updates. It is a positive real number (e.g., 0.01). A larger η accelerates convergence and may cause a single pit to miss the optimal solution. A smaller η is more stable but slows convergence. θ represents the gradient (first-order partial derivative vector) of the loss function L(θ) with respect to the parameter θ, pointing in the direction of fastest growth of the loss function; the negative gradient indicates a decreasing loss direction. In A / B testing, L(θ) can be a replica of the target metric (user retention rate, click-through rate), so the optimization goal is to maximize the metric. θ is the policy parameter, i.e., the variable to be optimized (such as the recommendation algorithm weights, ad bidding parameters, etc.). L(θ) represents the loss function, which is calculated by combining CVR and the ROI indicator in Table 6 using the following formula:

[0334] L(θ)=-α·CVR+β·ROI (6);

[0335] Among them, α and β are weight coefficients, which are used to balance the importance of CVR and ROI in the optimization objectives. They need to be adjusted according to business needs. The larger α is, the more the model tends to improve the conversion rate (possibly at the expense of ROI). The larger β is, the more the model tends to improve ROI (possibly at the expense of conversion rate).

[0336] 3. Full release of winning strategies:

[0337] Stop the test when all of the following conditions are met:

[0338] Statistical power (Power)>80%.

[0339] The minimum detectable effect (MDE) reached 5%.

[0340] The indicator was significantly positive for three consecutive days (p<0.05).

[0341] The health retail supply chain optimization method provided in this embodiment realizes the intelligent upgrade of the health retail supply chain, improves product selection accuracy, recommendation matching and marketing efficiency; deeply explores user needs, provides personalized health products and services, and improves user experience and satisfaction; optimizes resource allocation, reduces operating costs, and improves corporate profitability.

[0342] As one or more specific application embodiments of the present invention, combined with Figure 4 The health retail supply chain optimization method provided by the present invention is further described in detail. The specific process is as follows:

[0343] Step S1, obtain multi-source heterogeneous data sources through the data layer, including short video data (such as user viewing time, number of likes), historical order data (such as purchase records, return records) and external data (such as market trends, competitor prices) and other user behavior data and topic text data, and send the above data to the processing layer for processing.

[0344] In step S2, the processing layer uses data cleaning tools (such as Python's Pandas library) to deduplicate, fill in missing values, and standardize user behavior data such as short video data (such as user viewing time and number of likes), historical order data (such as purchase records and return records), and external data (such as market trends and competitor prices), as well as topic text data.

[0345] In step S3, the processing layer extracts consumer demand characteristics from user behavior data such as pre-processed short video data (such as user viewing time and number of likes), historical order data (such as purchase records and return records), and external data (such as market trends and competitor prices). At the same time, the processing layer performs multi-level semantic analysis on the text data, extracts explicit demand keywords and implicit demand keywords, and obtains user basic attribute characteristics.

[0346] In step S4, the processing layer constructs a user profile based on basic attribute characteristics, health demand characteristics, and consumption demand characteristics, and associates the preset product knowledge graph to obtain a product selection pool. The processing layer transmits the user profile, product selection pool, and all related data to the AI ​​model layer.

[0347] In step S5, the AI ​​model layer builds a multimodal large model based on user portraits, product selection pools, and all related data. The multimodal large model includes a demand forecasting model, a recommendation model, and a pricing model. The AI ​​model layer transmits the demand forecasting model, the recommendation model, and the pricing model to the application layer for actual application.

[0348] The processing steps of the demand forecasting model are optimization steps within the enterprise and are embedded in the enterprise ERP (Enterprise Resource Planning) system. The specific details correspond to the above-mentioned step S304 and will not be repeated here.

[0349] The recommendation model and pricing model are part of the enterprise's marketing steps for users, and both can be formed into APPs or mini-programs for users to use. The specific details of the recommendation model correspond to the above step S305, and the specific details of the pricing model correspond to the above step S306, which will not be repeated here.

[0350] The health retail supply chain optimization method provided in this embodiment integrates short video data, topic text data, historical order data and external data, uses a large model to explore users' potential needs, optimizes the product selection pool, and combines user portraits and price sensitivity analysis to achieve accurate recommendations and personalized marketing, thereby improving supply chain efficiency, user conversion rate and satisfaction.

[0351] In this embodiment, a health retail supply chain optimization device is also provided, which is used to implement the above-mentioned embodiments and preferred embodiments. The details that have been described will not be repeated here. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware, is also possible and conceivable.

[0352] This embodiment provides a health retail supply chain optimization device, such as Figure 6 Shown, including:

[0353] The data acquisition module 601 is used to acquire text data and user behavior data of health retail scenarios.

[0354] The keyword extraction module 602 is used to perform multi-level semantic analysis on text data using natural language processing technology to extract keywords of potential user needs.

[0355] The user portrait construction module 603 is used to construct a user portrait based on the user's potential demand keywords and the user behavior data.

[0356] The user grouping and product selection pool adjustment module 604 is used to group users based on preset grouping standards and user portraits in the health scenario, screen out corresponding health products for different user groups, and optimize and adjust the health products in demand based on the market trend data in the user behavior data to obtain an adjusted product selection pool.

[0357] The recommendation list generation module 605 is used to calculate the matching degree between the user and the preset candidate social platforms based on the user portrait and the product selection pool, and generate a social platform recommendation list based on the matching degree.

[0358] In some optional implementations, the keyword extraction module 602 includes:

[0359] The text data feature extraction unit is used to use natural language processing technology to perform sequence labeling, sentiment polarity analysis and topic distribution extraction on text data, and correspondingly identify user intentions, health topic attitudes and topic distribution.

[0360] The keyword generation unit is used to generate explicit demand keywords and implicit demand keywords based on user intention, health topic attitude, topic distribution and preset health knowledge graph.

[0361] In some optional implementations, the user behavior data includes social platform short video data, historical order data, social platform interaction data, and external data; the user portrait construction module 603 includes:

[0362] The basic attribute characteristics and health demand characteristics acquisition unit is used to obtain the user's basic attribute characteristics and extract health demand characteristics based on the user's potential demand keywords.

[0363] The consumer demand feature extraction unit is used to extract consumer demand features based on social platform short video data, historical order data, social platform interaction data and external data.

[0364] The user portrait construction unit is used to construct user portraits based on basic attribute characteristics, health demand characteristics and consumption demand characteristics.

[0365] In some optional implementations, the preset grouping criteria include health need characteristic criteria, consumption behavior characteristic criteria, health status characteristic criteria, and spatiotemporal characteristic criteria; the user grouping and product selection pool adjustment module 604 includes:

[0366] The user clustering unit is used to cluster users based on health demand feature standards, consumption behavior feature standards, health status feature standards, spatiotemporal feature standards and user portraits using a clustering algorithm to obtain different user groups.

[0367] The trend prediction result generating unit is used to extract trend feature vectors from market trend data, and input the trend feature vectors into a preset time series prediction model for training to obtain trend prediction results.

[0368] The weight adjustment unit is used to adjust the selection weight of in-demand health products based on trend prediction results and preset product selection decision rules.

[0369] The product selection pool determination unit is used to dynamically update and remove in-demand health products based on the adjusted weights to obtain an adjusted product selection pool.

[0370] In some optional implementations, the recommendation list generation module 605 includes:

[0371] A dense vector generation unit, used to convert user profiles into dense vectors using embedding technology;

[0372] A platform feature extraction unit is used to extract corresponding platform features of preset candidate social platforms and represent the platform features using vectorization; platform features include platform content, platform products, and real-time performance;

[0373] A matching degree calculation unit is used to calculate multiple matching degrees between dense vectors and corresponding platform feature vectors using a hybrid collaborative filtering algorithm based on the product selection pool;

[0374] The recommendation list generating unit is used to arrange the multiple matching degrees in descending order, and generate a social platform recommendation list based on the matching degrees arranged in descending order and the social platform operation strategy.

[0375] The recommendation list adjustment unit is used to obtain the real-time behavior of users on the social platform and dynamically adjust the social platform recommendation list based on the real-time behavior of users.

[0376] In some optional implementations, the health retail supply chain optimization device further includes:

[0377] The coupon formulation and optimization module is used to obtain users' historical consumption behavior and price sensitivity. In the process of generating recommendation lists on social platforms, operations optimization theory is used to formulate personalized coupon issuance strategies based on historical consumption behavior, price sensitivity and real-time user behavior. A / B testing algorithms are used to optimize personalized coupon issuance strategies.

[0378] The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.

[0379] The health retail supply chain optimization device in this embodiment is presented in the form of a functional unit, where the unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that executes one or more software or fixed programs, and / or other devices that can provide the above functions.

[0380] The embodiment of the present invention also provides a computer device having the above Figure 6 The health retail supply chain optimization device shown.

[0381] See also Figure 7 , Figure 7 is a structural diagram of a computer device provided by an optional embodiment of the present invention, such as Figure 7 As shown, the computer device includes: one or more processors 10, memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components utilize different buses to communicate with each other and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in the memory or on the memory to display the graphical information of the GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Equally, multiple computer devices can be connected, and each device provides part of the necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 7 A processor 10 is taken as an example.

[0382] The processor 10 may be a central processing unit, a network processor, or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit, a programmable logic device, or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic, or any combination thereof.

[0383] The memory 20 stores instructions that can be executed by at least one processor 10, so that the at least one processor 10 executes the method shown in the above embodiment.

[0384] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and application programs required for at least one function; the data storage area may store data created based on the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0385] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0386] The computer device further includes an input device 30 and an output device 40. The processor 10, the memory 20, the input device 30 and the output device 40 may be connected via a bus or other means. Figure 7 The bus connection is taken as an example.

[0387] The input device 30 can receive input digital or character information and generate key signal input related to user settings and function control of the computer device, such as a touch screen, a keypad, a mouse, a trackpad, a touch pad, an indicator stick, one or more mouse buttons, a trackball, a joystick, etc. The output device 40 can include a display device, an auxiliary lighting device (e.g., an LED), and a tactile feedback device (e.g., a vibration motor). The above-mentioned display device includes but is not limited to a liquid crystal display, a light emitting diode, a display, and a plasma display. In some optional embodiments, the display device can be a touch screen.

[0388] The embodiment of the present invention also provides a computer-readable storage medium. The above-mentioned method according to the embodiment of the present invention can be implemented in hardware, firmware, or implemented as a computer code that can be recorded in a storage medium, or implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and downloaded through a network and will be stored in a local storage medium, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state drive, etc.; further, the storage medium can also include a combination of the above-mentioned types of memory. It can be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor or hardware, the method shown in the above embodiment is implemented.

[0389] A portion of the present invention may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can call or provide the method and / or technical solution according to the present invention through the operation of the computer. Those skilled in the art should understand that the form in which the computer program instruction exists in a computer-readable medium includes, but is not limited to, a source file, an executable file, an installation package file, etc. Accordingly, the way in which the computer program instruction is executed by the computer includes, but is not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium may be any available computer-readable storage medium or communication medium that can be accessed by the computer.

[0390] Although the embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention. Such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A health retail supply chain optimization method, characterized in that: The method comprises: Obtain text data and user behavior data for health retail scenarios; Using natural language processing technology to perform multi-level semantic analysis on the text data to extract keywords of potential user needs; Building a user profile based on the user's potential demand keywords and the user behavior data; Users are grouped based on preset grouping criteria and user portraits in health scenarios, and corresponding health products are screened for different user groups. Furthermore, these products are optimized and adjusted based on market trend data from user behavior data to obtain an adjusted product pool. The matching degree between the user and the preset candidate social platforms is calculated based on the user portrait and the product selection pool, and a social platform recommendation list is generated based on the matching degree.

2. The method according to claim 1, characterized in that The natural language processing technology is used to perform multi-level semantic analysis on the text data to extract keywords of potential user needs, including: Using natural language processing technology to perform sequence labeling, sentiment polarity analysis, and topic distribution extraction on the text data, respectively, to identify user intentions, health topic attitudes, and topic distribution; Generate explicit demand keywords and implicit demand keywords based on user intention, attitude towards health topics, topic distribution and preset health knowledge graph.

3. The method according to claim 1, characterized in that The user behavior data includes social platform short video data, historical order data, social platform interaction data and external data; The step of constructing a user profile based on the user's potential demand keywords and the user behavior data includes: Obtaining basic attribute characteristics of the user, and extracting health demand characteristics based on the user's potential demand keywords; Extract consumer demand characteristics based on social platform short video data, historical order data, social platform interaction data and external data; A user profile is constructed based on the basic attribute characteristics, health demand characteristics and consumption demand characteristics.

4. The method according to claim 1, wherein The preset grouping standards include health demand characteristic standards, consumption behavior characteristic standards, health status characteristic standards and time and space characteristic standards; The user grouping based on preset grouping criteria and user portraits in the health scenario includes: Based on health demand characteristic standards, consumption behavior characteristic standards, health status characteristic standards, spatiotemporal characteristic standards and user portraits, a clustering algorithm is used to group users to obtain different user groups.

5. The method according to claim 1, wherein The aforementioned optimization and adjustment of the health products in demand by combining the market trend data in the user behavior data to obtain the adjusted product pool includes: Extracting trend feature vectors from market trend data, and inputting the trend feature vectors into a preset time series prediction model for training to obtain trend prediction results; Adjust the selection weights of in-demand health products based on trend forecast results and preset selection decision rules; Based on the adjusted weights, in-demand health products are dynamically updated and removed from the shelves to obtain an adjusted product selection pool.

6. The method according to claim 1, characterized in that The calculating the matching degree between the user and the preset candidate social platforms based on the user portrait and the product selection pool, and generating a social platform recommendation list based on the matching degree, includes: Using embedding technology to convert the user profile into a dense vector; Extracting corresponding platform features of preset candidate social platforms and representing the platform features using vectorization; the platform features include platform content, platform products, and real-time performance; Based on the product selection pool, a hybrid collaborative filtering algorithm is used to calculate multiple matching degrees between dense vectors and corresponding platform feature vectors; Arrange multiple matching degrees in descending order, and generate a social platform recommendation list based on the matching degrees after descending order and the social platform operation strategy.

7. The method according to claim 6, characterized in that Generate a social platform recommendation list based on the matching degree and social platform operation strategy in descending order, including: The real-time behavior of users on the social platform is obtained, and the social platform recommendation list is dynamically adjusted based on the real-time behavior of users.

8. The method according to claim 7, characterized in that The method further comprises: Obtain users' historical consumption behavior and price sensitivity; In the process of generating the recommendation list on the social platform, based on the historical consumption behavior, price sensitivity and real-time user behavior, operational optimization theory is used to formulate a personalized coupon distribution strategy; An A / B testing algorithm is used to optimize the personalized coupon distribution strategy.

9. A health retail supply chain optimization device, characterized in that: The device comprises: Data acquisition module, used to obtain text data and user behavior data of health retail scenarios; A keyword extraction module is used to perform multi-level semantic analysis on the text data using natural language processing technology to extract keywords of potential user needs; A user portrait construction module is used to construct a user portrait based on the user's potential demand keywords and the user behavior data; The user grouping and product selection pool adjustment module is used to group users based on preset grouping criteria and user portraits in health scenarios, screen out corresponding health products for different user groups, and optimize and adjust the health products in demand based on market trend data in user behavior data to obtain an adjusted product selection pool; The recommendation list generation module is used to calculate the matching degree between the user and the preset candidate social platforms based on the user portrait and the product selection pool, and generate a social platform recommendation list based on the matching degree.

10. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the health retail supply chain optimization method according to any one of claims 1 to 8 by executing the computer instructions.

Images