Marketing strategy optimization management system based on six elements of order transaction

Abstract

The invention discloses a marketing strategy optimization management system based on six elements of order transaction, and belongs to the field of communication management systems, in terms of data acquisition, multi-source comprehensive information collection enables enterprises to perceive consumer demands in all directions and no longer blindly grope, a dynamic feature modeling unit fuses various kinds of data into six-dimensional feature vectors, and the six-dimensional feature vectors are integrated into a database; multiple factors of commodities, users and festivals are balanced and considered, a solid foundation is laid for a marketing strategy, a commodity-festival-user ternary association graph constructed by a festival graph generation engine enables commodities to be pushed in a targeted manner, a personalized strategy is generated by an intelligent decision module, the matching degree of the commodities and consumers is improved, and the marketing efficiency is improved. The overall continuous feedback optimization mechanism of the system enables the marketing strategy to be like a continuously evolved life entity, can better adapt to the market change, and improves the sales probability of commodities.

Description

Technical Field

[0001] The present invention relates to the field of communication management systems, and more specifically, to a marketing strategy optimization management system based on six elements of order transactions. Background Art

[0002] In today's fiercely competitive business environment, holiday marketing has become a crucial tool for companies to boost sales and enhance brand influence. Consumers' shopping needs and preferences shift significantly during various holidays, and companies hope to accurately capitalize on these changes and recommend appropriate products to consumers, thereby increasing the likelihood of sales.

[0003] Traditional marketing strategies often have many limitations when dealing with holiday marketing. On the one hand, data collection is not comprehensive and timely enough. Many companies only rely on certain channels to obtain user information and market data. For example, they only focus on users' historical purchase records, but ignore the multi-dimensional information of users' behavior on social media and their emotional tendencies towards different festivals. This is like groping in the dark, only seeing limited light, and it is difficult to fully understand the real needs of consumers during the holidays. On the other hand, there is a lack of effective data analysis and modeling methods. Although the existing marketing system can perform simple data statistics and analysis, it lacks the ability to deeply explore the complex relationships between products, festivals and users. It is like a puzzle. You can only see scattered small pieces, but it is difficult to put them together into a complete picture. It is difficult to build a precise product push map suitable for different festivals.

[0004] In addition, the current existing system has a poor effect in comprehensively considering the correlation of multiple key factors in order transactions, such as product value, user preferences, holiday attributes, market trends, supply chain status and promotion strategies. This is like a battle without overall planning, with a lack of coordination and cooperation among various combat links, making it difficult to achieve ideal results.

[0005] To sum up, the existing marketing systems and strategies have many shortcomings when dealing with holiday marketing. There is an urgent need for an innovative technical solution that can comprehensively and accurately collect data, deeply analyze the correlations behind the data, and flexibly adjust marketing strategies to build item push maps suitable for different holidays and increase the probability of successful sales of goods on specific holidays. Therefore, we proposed a marketing strategy optimization management system based on the six elements of order transaction to solve the above problems. Summary of the Invention

[0006] 1. Technical problems to be solved

[0007] In response to the problems existing in the prior art, the purpose of the present invention is to provide a marketing strategy optimization management system based on the six elements of order transaction. In terms of data collection, multi-source comprehensive information collection allows enterprises to fully perceive consumer needs and no longer grope blindly. The dynamic feature modeling unit integrates various types of data into a six-dimensional feature vector, balancing the multiple factors of goods, users, and festivals to lay a solid foundation for marketing strategies. The product-festival-user ternary association map constructed by the festival map generation engine allows product push to be targeted. The personalized strategy generated by the intelligent decision-making module improves the matching degree between goods and consumers. The overall continuous feedback optimization mechanism of the system makes the marketing strategy like a constantly evolving living organism, which can better adapt to market changes and increase the probability of product sales.

[0008] 2. Technical solution

[0009] To solve the above problems, the present invention adopts the following technical solutions.

[0010] A marketing strategy optimization management system based on the six elements of order completion, including a multi-source data acquisition module, a dynamic feature modeling unit, a holiday map generation engine, and an intelligent decision-making module;

[0011] The multi-source data acquisition module includes:

[0012] Behavioral data collection sub-unit: This unit deploys tracking code on shopping platform pages and social media interactive interfaces to capture user browsing history, search keywords, product collections, and add-to-cart behavior logs in real time.

[0013] Order data interface subunit: connects to the enterprise order management system through a preset data interaction protocol and periodically extracts structured data of historical orders;

[0014] Product data extraction subunit: establishes a connection with the product management system's database and extracts product attribute information through data synchronization tools;

[0015] Festival data crawling sub-unit: uses a distributed crawler framework to crawl festival-related data from designated network channels, and has a built-in data cleaning component to filter out invalid information;

[0016] The dynamic feature modeling unit includes:

[0017] Original feature extractor: performs field analysis on the behavior logs and order data output by the multi-source data acquisition module to extract six types of original features such as product value and user preferences;

[0018] Feature fusion processor: normalizes and fuses the six types of original features through a preset weight matrix;

[0019] Vector generator: maps the fused features into six-dimensional feature vectors, where the six-dimensional feature vectors correspond to product value, user preference, holiday attributes, market trends, supply chain status, and promotion strategies respectively.

[0020] The festival map generation engine includes:

[0021] Triple recognition component: uses an entity recognition algorithm to extract product ID, holiday type, and user ID entities from the six-dimensional feature vector, and uses a relationship extraction model to identify the association between the three;

[0022] Graph storage component: uses a graph database to store ternary relationships and constructs a heterogeneous network of products, festivals, and users using a node-edge structure.

[0023] Dynamic update component: sets a time window, repeats the triplet recognition process for new data, and adjusts the weights of edges in the graph through an incremental update mechanism;

[0024] The intelligent decision-making module includes:

[0025] Strategy Generator: Based on the node association weights of the festival graph, combined with collaborative filtering algorithms and graph attention mechanisms, it generates product combination plans, price adjustment coefficients, and channel placement priorities.

[0026] Feedback Optimizer: Receives user response data after strategy execution, adjusts the parameters of the strategy generator using a gradient descent algorithm, and synchronizes the optimization results to the dynamic update component of the festival map generation engine;

[0027] Each subunit of the multi-source data acquisition module pushes data to the original feature extractor of the dynamic feature modeling unit through the HTTP protocol, the vector generator of the dynamic feature modeling unit transmits the six-dimensional feature vector to the triple recognition component of the festival map generation engine through the RPC interface, the map storage component of the festival map generation engine provides map data to the strategy generator of the intelligent decision-making module through the data subscription mechanism, and the feedback optimizer of the intelligent decision-making module feeds back strategy effect data to the dynamic update component of the festival map generation engine in real time through WebSocket.

[0028] Furthermore, the historical order data fields extracted by the order data interface subunit include order number, order timestamp, product SKU, purchase quantity, actual payment amount, payment method, and delivery address latitude and longitude;

[0029] The commodity attribute data fields obtained by the commodity data extraction subunit include commodity ID, name, specification parameters, material code, color code, price, real-time inventory, brand code and category;

[0030] The processing flow of the festival data crawling subunit is:

[0031] The date, holiday arrangement and folk activity description text of statutory holidays are crawled from news websites and government announcement pages;

[0032] Festival traditional consumption habit data is scraped from folk culture websites;

[0033] Festival-related text is crawled from social media platforms and transmitted to the sentiment analysis subunit, which uses a pre-trained BERT model for sentiment classification. The classification output probability includes positive, negative and neutral probability;

[0034] Sales data in the same period is extracted from the enterprise historical database and transmitted to the time series analysis subunit, which uses ARIMA and SARIMA models to fit the data and outputs trend prediction parameters, including mean, variance and periodic fluctuation coefficient.

[0035] Further, the original feature extractor of the dynamic feature modeling unit extracts six types of features through the following technical means:

[0036] Commodity value features: Calculate three core indicators, gross profit margin = (sales price - cost price) / sales price, inventory turnover rate = sales in the past 30 days / average inventory, and sales growth rate = (this month's sales - last month's sales) / last month's sales. The comprehensive score is obtained by weighted summation;

[0037] User preference features: Based on user behavior data and historical order data, a collaborative filtering algorithm is used to generate user preference for product categories. The collaborative filtering algorithm is based on user-item rating matrix, and the similarity is calculated using cosine distance. The deep learning model outputs user price sensitivity and brand loyalty. The user price sensitivity ranges from 0 to 1, and the higher the value, the more sensitive to price. Brand loyalty is calculated based on the frequency of repeated purchases of the same brand;

[0038] Festival attribute features: The festival is divided into "statutory holiday", "folk festival" and "seasonal node" types by classification algorithm, and the sales impact coefficient of each product category is calculated based on historical sales data;

[0039] Market trend features: Connect to industry database API to obtain competitor prices and promotion activity data, combined with macroeconomic indicators, use LSTM model to predict market demand growth rate and price fluctuation range in the next 30 days;

[0040] Supply chain status features: Through real-time interface with inventory management system and logistics system, obtain supply stability score, replenishment cycle and logistics delivery time.

[0041] Promotion strategy features: Extract historical promotion methods from the company's marketing activity database, compare the conversion rate improvement and average order value change rate of different methods through A / B testing, and generate applicable product categories and optimal discount ranges for each promotion method.

[0042] Furthermore, in the triple recognition component of the festival graph generation engine, the product-festival association calculation process based on the graph neural network is as follows:

[0043] Construct a heterogeneous information network: use the node attributes of products, festivals, and users as nodes, and "user purchases products" and "product adaptation to festivals" as edges. The node attributes of the products include value ratings and categories, the node attributes of festivals include types and influence coefficients, and the node attributes of users include preference feature vectors. The edge weight calculation formula for "user purchases products" is: edge weight = purchase frequency / total consumption times. The edge weight calculation formula for "product adaptation to festivals" is: edge weight = sales during festivals / sales during non-festivals.

[0044] The network is trained using a graph convolutional network: the node feature vector is input, the node embedding representation is learned through two layers of convolutional layers, and the cosine similarity of the product-festival node pair is output as the association score.

[0045] Furthermore, in the strategy generator of the intelligent decision-making module, the user portrait and festival feature matching model process integrating the attention mechanism is as follows:

[0046] Sort user historical behavior data by timestamp, use LSTM network to embed time series, and obtain user behavior time series feature vector;

[0047] Festival feature processing: Festival sentiment, influence coefficient, and product attribute features are combined into a comprehensive feature vector through feature splicing;

[0048] Multi-head attention mechanism processing: Set 8 attention heads to learn the association weights of temporal features and comprehensive features in different subspaces, and generate a recommendation vector after weighted summation.

[0049] Furthermore, in the feedback optimizer of the intelligent decision-making module, the push strategy dynamic optimization algorithm process based on reinforcement learning is as follows:

[0050] Define the state space: S = product node characteristics, holiday node characteristics, user node characteristics, and the action space: A = push product combination, price adjustment value, channel selection;

[0051] Reward function: R = 0.3 × click volume + 0.5 × conversion rate + 0.2 × average order value;

[0052] Adopting the Actor-Critic framework: the Actor network generates the action probability distribution, and the Critic network evaluates the state-action value;

[0053] Prioritized experience replay mechanism: Sort the S, A, R, and S' experiences by |R|, prioritize sampling high-reward experiences to update network parameters, and iterate once every 1,000 new data points.

[0054] The marketing strategy optimization management method based on the six elements of order closing includes the following steps:

[0055] S1. Build a knowledge base of six festival characteristics:

[0056] S11: Each subunit of the multi-source data collection module collects data in parallel. After removing outliers through the data cleaning component, the data is stored in the data lake according to four categories: "user behavior", "historical orders", "product attributes", and "holiday characteristics".

[0057] S12. The original feature extractor of the dynamic feature modeling unit reads data from the data lake and extracts six types of original features, which are standardized into a unified format;

[0058] S13. Store the standardized features into the knowledge base according to the association key of "product ID-holiday ID-user ID":

[0059] S2. Training the product-festival association model:

[0060] S21. Extract historical order data from the knowledge base and build a heterogeneous information network;

[0061] S22. Define meta-paths: "product-user-festival" and "festival-product-user". Use the random walk algorithm to sample paths in the network and extract product-festival co-occurrence patterns.

[0062] S23. Use graph attention mechanism to calculate node importance: for each node, aggregate neighbor node features weighted by attention coefficient, and optimize model parameters through back propagation;

[0063] S3. Generate personalized recommendation list:

[0064] S31. Sort the target user's historical behavior data by time and extract temporal features using an LSTM network.

[0065] S32, extracting the emotional characteristics and influence coefficient of the current festival and the attribute characteristics of the candidate products, and integrating them into comprehensive characteristics;

[0066] S33. Calculate the matching degree between user temporal features and product comprehensive features through a multi-head attention mechanism, and generate a top 30 recommendation list in descending order of matching degree.

[0067] S4. Online optimization push strategy:

[0068] S41. Push the recommendation list to the user contact channel and collect user response data in real time;

[0069] S42, the feedback optimizer converts the response data into a reward value R, and the reinforcement learning algorithm updates the policy generator parameters;

[0070] S43. Synchronize the optimized product-festival association weights to the festival graph generation engine, update the weights of the corresponding edges in the graph, and achieve closed-loop optimization of the strategy.

[0071] 3. Beneficial effects

[0072] Compared with the prior art, the advantages of the present invention are:

[0073] (1) In this solution, data collection is more comprehensive. The multi-source data collection module can obtain multi-dimensional data on user behavior, historical orders, product attributes, and holiday characteristics in real time and comprehensively. This allows companies to no longer rely on one-sided information to make marketing decisions as in the past. Instead, it is like having a detailed map, which allows them to clearly understand the overall market situation and consumer needs, laying a solid foundation for subsequent precision marketing. Data processing is more accurate. The dynamic feature modeling unit constructs a six-dimensional feature vector based on the collected data, including product value, user preferences, holiday attributes, market trends, supply chain status, and promotion strategies. This accurate data processing method allows companies to gain a deep insight into the intrinsic connection between products, markets, and users, and reduce the probability of marketing errors caused by inaccurate information.

[0074] (2) In this solution, the push map is more accurate. The holiday map generation engine uses knowledge graph technology to construct a ternary relationship between products, festivals and users, forming a dynamically updated holiday product push map. It can accurately push suitable products according to different holiday scenes and user needs, greatly improving the pertinence and effectiveness of product push and reducing the waste of ineffective marketing resources; the decision-making strategy is more flexible, and the intelligent decision-making module generates personalized push strategies based on the push map, including product combination plans, price adjustment suggestions and channel delivery plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 Schematic diagram of the system architecture of the present invention;

[0076] Figure 2 This is a schematic diagram of the holiday data processing process principle of the present invention;

[0077] Figure 3 Schematic diagram of the principle of the six-category feature extraction method of the present invention;

[0078] Figure 4Schematic diagram of the heterogeneous network construction and correlation calculation principle of the present invention;

[0079] Figure 5 This is a schematic diagram of the user portrait and festival feature matching model principle of the present invention;

[0080] Figure 6 This is a schematic diagram of the principle of the push strategy dynamic optimization algorithm of the present invention;

[0081] Figure 7 Schematic diagram of the method steps of the present invention. DETAILED DESCRIPTION

[0082] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the specification of the present invention; it is obvious that the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0083] Example 1:

[0084] See also Figure 1-Figure 7 The marketing strategy optimization management system based on the six elements of order completion includes a multi-source data acquisition module 1, a dynamic feature modeling unit 2, a holiday map generation engine 3, and an intelligent decision-making module 4. The specific principles are as follows:

[0085] 1. Multi-source data collection: building a global data foundation

[0086] The multi-source data acquisition module 1 operates in parallel through four sub-units to achieve comprehensive capture of global data, providing raw materials for subsequent analysis. Specific data conversion and transmission rules are followed during the acquisition process:

[0087] Behavioral data collection subunit 101: Deploy tracking code on shopping platform pages and social media interactive interfaces. Every user action (such as clicking, sliding, and inputting) will trigger the code to record relevant information, forming unstructured log data. After this data is cached locally, it is batch-transmitted to the data preprocessing server in JSON format via HTTP POST requests at preset time intervals. During the transmission process, the MD5 encryption algorithm is used to verify the data to ensure its integrity and accuracy. The receiving end determines whether the data is complete by comparing the MD5 value.

[0088] Order data interface subunit 102: Establishes a stable connection with the enterprise order management system based on a pre-established data interaction protocol and periodically initiates data requests. After the order management system responds to the request, it returns structured data that meets the query criteria (such as orders within a specific time period) encapsulated in XML or JSON format in an HTTP response body. After receiving the data, the order data interface subunit performs preliminary data cleaning to remove duplicate and invalid order records, and then stores the cleaned data in a local relational database temporary table for further processing.

[0089] Product data extraction subunit 103: Uses a data synchronization tool to establish a real-time connection with the product management system database. Using the database's CDC technology, it monitors changes in product attribute information tables in real time. Once data is updated (such as product listings, price adjustments, and inventory changes), the CDC mechanism captures these changes and transmits the changed data incrementally via the JDBC interface to the product data extraction subunit. After receiving the data, the subunit converts and integrates it according to the preset data format specifications to ensure data consistency, and then stores it in the corresponding product data directory in the local distributed file system.

[0090] Festival data crawling sub-unit 104: Utilizing a distributed crawler framework, the system acquires festival-related data from multiple channels. When crawling statutory holiday dates, adjustment schedules, and descriptions of folk activities from news websites and government announcement pages, the crawler simulates browser behavior, sending HTTP GET requests to retrieve webpage content. It then uses an HTML parsing library to extract the required information and converts it into structured JSON formatted data. When crawling festival-related text from social media platforms, the system calls the social media's open API to obtain text data related to specified keywords (such as festival names). Natural language processing tools are then used to perform part-of-speech tagging and stemming preprocessing, and the processed data is stored alongside the source platform information. When extracting contemporaneous sales data from a company's historical database, the system uses SQL queries to obtain data, uses data conversion tools to convert it into a format suitable for analysis, and transfers it to a local storage device via FTP. After acquiring the data, the system uses a built-in cleaning component to filter out invalid information based on regular expressions and data dictionary rules, retaining valid data for subsequent analysis.

[0091] 2. Dynamic Feature Modeling: Extracting the Six Elements of Order Transaction Features

[0092] The dynamic feature modeling unit 2 performs in-depth processing on the collected multi-source data, converting the raw data into a quantifiable "six-element feature vector of order completion" and clearly defining the data conversion and transmission process at each link:

[0093] Raw feature extraction: After the raw feature extractor 201 obtains data from different data sources, it extracts features using multi-dimensional algorithms and transmits the feature data to the subsequent processing link in a timely manner after extraction;

[0094] Product value characteristics: The product data stored in the distributed file system by the product data extraction subunit, the sales price, cost price, 30-day sales volume, and average inventory information of the product are read, three indicators are calculated, i.e., "gross profit rate = (sales price - cost price) / sales price", "inventory turnover rate = 30-day sales volume / average inventory", and "sales growth rate = (this month's sales - last month's sales) / last month's sales", and then weighted summation is performed according to the pre-set weights (such as gross profit rate weight 0.4, inventory turnover rate weight 0.3, and sales growth rate weight 0.3) to generate a comprehensive score. After the calculation is completed, the product value characteristic data is transmitted to the feature fusion processor 202 through a memory sharing mechanism (such as a shared memory queue);

[0095] User preference characteristics: The user behavior log data transmitted by the behavior data collection subunit and the historical order data provided by the order data interface subunit are imported into the user preference analysis module. Based on the user-product score matrix, the collaborative filtering algorithm for calculating similarity through cosine distance generates category preference degree, and a deep learning model (such as a multi-layer perceptron) is used to train the user historical purchase behavior data to output price sensitivity (the higher the value, the more sensitive) and brand loyalty based on repeat purchase frequency in the 0-1 interval. These user preference characteristic data are transmitted to the feature fusion processor 202 in binary serialized format through network sockets to reduce data transmission volume and transmission time;

[0096] Holiday attribute characteristics: The holiday-related structured data obtained by the holiday data crawling subunit is loaded into the holiday attribute analysis module. The holidays are divided into "statutory holidays", "folk festivals", and "seasonal nodes" through a classification algorithm (such as a decision tree algorithm), and the historical sales data obtained from the order data interface subunit are combined to calculate the sales influence coefficient of each holiday type on each product category. The calculation results are in the form of key-value pairs (such as {"Spring Festival - food category": 1.5}), which are transmitted to the feature fusion processor 202 through a lightweight message queue to ensure the reliability and asynchrony of data transmission;

[0097] Market trend characteristics: Competitor data are obtained by connecting to industry database APIs, and macroeconomic indicator data are obtained from a macroeconomic data platform. After integrating these data, they are input into an LSTM model for training and prediction. The market trend characteristic data predicting the market demand growth rate and price fluctuation range in the next 30 days are packaged in JSON format and transmitted to the feature fusion processor 202 through an HTTP PUT request for easy reception and processing;

[0098] Supply chain status features: Establish real-time interfaces with inventory and logistics systems to obtain data on supply stability scores, replenishment cycles, and logistics delivery timeliness. After undergoing simple local data format conversion (e.g., unifying the time format), these data are transmitted to the feature fusion processor 202 via RPC (framework) in a synchronous call manner to ensure data timeliness and accuracy.

[0099] Promotional strategy features: Extract historical promotional activity data from the marketing activity database, use A / B testing to compare the conversion rate improvement and average order value change of different promotional methods (such as discounts, full-amount discounts, and gifts), and generate applicable categories and optimal discount ranges for each promotional method through data analysis and mining. Promotional strategy feature data is transmitted in tabular form to the feature fusion processor 202 via a file transfer protocol (such as SFTP, secure file transfer protocol) to ensure data transmission security;

[0100] Feature fusion and vector generation: The feature fusion processor 202 normalizes and fuses the six types of features based on a preset weight matrix. After receiving data from each original feature extraction module, it first unifies the format and converts the data to ensure that the data can correctly participate in the fusion calculation. For example, all feature data are converted to floating-point types, and the data range is normalized to the interval of 0-1. Then, weighted calculations are performed according to the weight matrix to obtain the fused feature data. The vector generator 203 maps the fused features into a six-dimensional vector, and each dimension corresponds to the above six types of features. The generated six-dimensional feature vector is provided to the festival map generation engine 3 in a shared memory manner through memory mapping file technology, which facilitates its rapid reading and use and reduces data transmission overhead.

[0101] 3. Festival Graph Generation: Building a Product-Festival-User Association Network

[0102] The Festival Graph Generation Engine 3 constructs a dynamically updated "product-festival-user" heterogeneous network through entity association analysis, intuitively presenting the strength of the association between the three, and ensuring smooth data conversion and transmission during the construction and update process:

[0103] Triple recognition: The triple recognition component 301 obtains a six-dimensional feature vector from the vector generator 203 of the dynamic feature modeling unit 2 in a shared memory manner. An entity recognition algorithm (BERT-CRF model based on named entity recognition) is used to extract the "product ID", "festival type" and "user ID" entities from the six-dimensional feature vector, and a relationship extraction model (relationship extraction model based on convolutional neural network) is used to identify the relationship between the three (such as "user A buys product B during the Spring Festival"). The identified triple data is stored in the local memory in the form of a linked list and is regularly transmitted to the graph storage component 302 through a memory sharing mechanism;

[0104] Graph storage and structure definition: The graph storage component 302 uses a graph database (such as Neo4j) to store association relationships in a "node-edge" structure. After receiving the triple data transmitted by the triple recognition component, the data is first parsed and verified to ensure the integrity and accuracy of the data. For the "product ID", "festival type" and "user ID" entities, corresponding nodes are created respectively, and the node attributes (such as product nodes containing value ratings and categories; festival nodes containing types and influence coefficients; user nodes containing preference feature vectors) are stored in the node's attribute table. For association relationships, corresponding edges are created, and the direction and weight of the edges are determined according to the relationship type in the triple (such as "user purchases product" edge weight = purchase frequency / total consumption times; "product adaptation festival" edge weight = sales during festivals / non-festival sales). After the storage is completed, a data storage success notification is sent to the dynamic update component 303 so that it can perform subsequent operations.

[0105] Dynamic update mechanism: The dynamic update component 303 sets a time window (such as daily / weekly) and repeats the triple recognition process for the newly added data. It obtains the newly collected data from the multi-source data acquisition module 1 and the new data generated after the policy execution feedback from the intelligent decision-making module 4 by subscribing to the message queue (the topic related to new data in Kafka). After obtaining the new data, the triple recognition component 301 is called to perform entity and relationship recognition, and the recognition results are transmitted to the graph storage component 302. The graph storage component 302 incrementally updates the existing graph based on the new data and adjusts the weight of the edge (for example, when a festival is approaching, the weight of the "commodity-festival" edge increases dynamically with the increase of consumption data). After the update is completed, the intelligent decision-making module 4 is notified that the graph has been updated by publishing a message to the message queue so that it can obtain the latest association relationship data for strategy generation.

[0106] 4. Intelligent Decision-making and Optimization: Generating Strategies and Dynamically Iterating

[0107] Intelligent Decision Module 4 generates personalized marketing strategies based on the associations of holiday graphs and combines multiple algorithms. It also implements closed-loop optimization through real-time feedback. Data conversion and transmission work closely together throughout the entire process.

[0108] Strategy generation: The strategy generator 401 obtains the "product-festival-user" heterogeneous network data from the graph storage component 302 of the festival graph generation engine 3, obtains the JSON format file of the graph data through the network protocol, and then parses it into a graph data structure in the memory. It integrates the collaborative filtering algorithm and the graph attention mechanism, and identifies highly correlated "product combinations" (such as the combination of Mid-Autumn Festival, moon cakes, and red wine) based on the association weights of the festival graph nodes; calculates the user's attention to different products through the graph attention mechanism, and generates a price adjustment coefficient (such as providing higher discounts to price-sensitive users); and combines user behavior data (such as active social media channels) to determine the channel delivery priority. The generated marketing strategy data (such as product combination recommendation list, price adjustment plan, channel delivery plan) is encapsulated in XML format and sent to the strategy execution module (marketing activity management system) through the message queue, which is responsible for the implementation of specific strategies.

[0109] Feedback optimization: The feedback optimizer 402 implements dynamic strategy iteration through reinforcement learning. It obtains feedback data after strategy execution from the strategy execution module (marketing campaign management system), such as user click volume, conversion rate, and average order value. These data are transmitted to the feedback optimizer 402 in JSON format through the network interface. The feedback optimizer 402 defines the state space (products, festivals, user node features), action space (push combination, price adjustment, channel selection), and reward function (R = 0.3 × click volume + 0.5 × conversion rate + 0.2 × average order value); using the Actor-Critic framework, the Actor network generates the action probability distribution, and the Critic network evaluates the state-action value; through the priority experience replay mechanism (sampling high-reward experience by sorting by |R|), the parameters are iterated once every 1,000 new data are received. The iteratively updated parameters are transmitted to the strategy generator 401 through the memory sharing mechanism. At the same time, the feedback optimizer 402 sends the strategy optimization results in the form of messages to the dynamic update component 303 of the festival map generation engine 3, so that it can update the weights of the corresponding edges in the map according to the optimized strategy effects, thereby realizing the closed loop of "strategy generation-effect feedback-parameter optimization".

[0110] 5. System Data Interaction Process

[0111] Each module implements data flow through standardized protocols to ensure efficient collaboration throughout the entire process:

[0112] After completing data collection and preliminary processing, the four subunits of the multi-source data collection module 1 transmit the data to the original feature extractor 201 of the dynamic feature modeling unit 2 according to their respective set transmission methods and time intervals. For example, the behavior data collection subunit transmits log data in JSON format via HTTP POST, and the order data interface subunit transmits order data in XML or JSON format via HTTP response;

[0113] The vector generator 203 of the dynamic feature modeling unit 2 transmits the generated six-dimensional feature vector to the triple recognition component 301 of the festival map generation engine 3 in a shared memory manner through the memory mapping file technology, thereby achieving efficient data sharing and transmission;

[0114] After completing the graph data storage, the graph storage component 302 of the festival graph generation engine 3 provides the graph data to the strategy generator 401 of the intelligent decision module 4 through message notification or subscription-publishing mechanism (based on message queue). The strategy generator 401 obtains the graph data through network request to generate strategy.

[0115] The feedback optimizer 402 of the intelligent decision-making module 4 feeds back the strategy effect data to the dynamic update component 303 of the festival map generation engine 3 in real time via WebSocket, so that the dynamic update component 303 updates the map according to the new data and transmits the optimized parameters to the strategy generator 401 for subsequent strategy generation;

[0116] The marketing strategy optimization management system based on the six elements of order transaction realizes the precise generation and dynamic adjustment of marketing strategies based on the six elements of order transaction through the whole process of "data collection-feature modeling-graph construction-decision optimization". It can continuously optimize product portfolio, price and channel strategies according to dynamic changes in user preferences, holiday attributes and market trends, thereby improving marketing conversion efficiency.

[0117] Example 2:

[0118] In view of the above embodiment 1, for further description, please refer to Figure 1-Figure 7 In order to enable those skilled in the art to more clearly understand the present invention, the following takes the Spring Festival promotion activities of a large e-commerce platform as an example, combines the working principles of each module of the system, and elaborates on the specific application process of the system in detail, fully reflecting the role of the six elements of order transaction in actual marketing scenarios.

[0119] 1. Data collection and processing before the event (45 days before the Spring Festival)

[0120] 1. Full operation of multi-source data acquisition module 1

[0121] Behavioral Data Collection Sub-Unit 101: Deploy tracking code on the platform's homepage, product details, New Year's shopping section, Spring Festival event pages, and other pages, as well as on the platform's official Weibo, WeChat official account, TikTok, and other social media interfaces. When users browse, the code records the user's dwell time on each page and the order in which they browse. When users search, it captures search keywords such as "Spring Festival gifts," "New Year's Eve dinner ingredients," and "New Year's clothes." For user product collection and purchase actions, a behavior log is also generated in real time, including information such as the collection and purchase time and product ID. This data is batch-transferred to the data preprocessing server every 5 minutes via HTTP POST requests in JSON format. The MD5 encryption algorithm is used for data verification during transmission to ensure data integrity and accuracy.

[0122] Order Data Interface Subunit 102: Establishes a stable connection with the enterprise's order management system according to a pre-set data exchange protocol. Every day at 3:00 AM, the system automatically initiates a data request to extract structured data from historical orders during the Spring Festival period over the past five years. This data includes fields such as order number, order timestamp, product SKU, quantity purchased, actual payment amount, payment method, and the latitude and longitude of the delivery address. This data is transmitted in XML format via a RESTful API interface. After cleaning to remove duplicate and abnormal orders, it is stored in the historical order table in the MySQL database.

[0123] Product data extraction subunit 103 uses a data synchronization tool to establish a connection with the product management system's database and extract product attribute information in real time. For items likely to be popular during the Spring Festival, such as food gift boxes, clothing, shoes, and hats, and household items, detailed data such as product ID, name, specifications, material code, color code, price, real-time inventory, brand code, and category are obtained. This data is transferred to the HDFS distributed file system via the JDBC interface and stored by product category for subsequent processing.

[0124] The Festival Data Crawling Sub-unit 104 uses a distributed crawler framework to crawl Spring Festival-related data from multiple designated online channels. It crawls news websites and government announcement pages to collect information on statutory Spring Festival holidays, work arrangements, and descriptions of local folk customs, such as the time and location of temple fairs. It also extracts data on traditional Spring Festival consumption habits, such as common New Year's Eve dinner dishes and New Year's gift preferences, from folk culture websites. It also crawls Spring Festival-related text from social media platforms and transmits it to the Sentiment Analysis Sub-unit, which uses a pre-trained BERT model to classify sentiment and output probabilities of positive, negative, and neutral. For example, "I'm very happy to buy New Year's goods during the Spring Festival" is classified as a positive sentiment with a probability of 0.92. Sales data for the same period during the Spring Festival over the past five years is extracted from the company's historical database and transmitted to the Time Series Analysis Sub-unit, which fits the data using ARIMA and SARIMA models and outputs trend prediction parameters, including mean, variance, and cyclical fluctuation coefficient, to provide a basis for subsequent sales forecasts. The crawled data is filtered out by a built-in data cleaning component and stored in a local database.

[0125] 2. Preliminary processing of dynamic feature modeling unit 2

[0126] Original feature extractor 201: Performs field parsing on the behavior logs, order data, and other data output by the multi-source data acquisition module 1. Six original features, such as product value and user preferences, are extracted from the behavior logs and order data. Taking the product value feature as an example, the gross profit margin of a food gift box is calculated as (sales price - cost price) / sales price, resulting in a gross profit margin of 35%. The inventory turnover rate over the past 30 days is calculated as sales volume over the past 30 days / average inventory, resulting in a value of 2.5. The sales growth rate is calculated as (this month's sales volume - last month's sales volume) / last month's sales volume, resulting in a value of 15%. Through weighted summation (weights are 0.4, 0.3, and 0.2, respectively), the comprehensive score is 0.35×0.4+2.5×0.3+0.15×0.2=0.14+0.75+0.03=0.92. For user preference characteristics, based on the user-product rating matrix, a collaborative filtering algorithm using cosine distance to calculate similarity is used to generate user preferences for food categories, clothing categories, etc. At the same time, the deep learning model outputs user price sensitivity. For example, user F's price sensitivity is 0.7, indicating that he is relatively sensitive to price. Brand loyalty is also calculated based on the frequency of repeated purchases of the same brand. For example, user G has purchased a certain brand of nut gift boxes many times, and his brand loyalty is 0.8.

[0127] Feature Fusion Processor 202: Normalizes and fuses the six original features using a preset weight matrix. The weight matrix is ​​set based on historical marketing data and expert experience. For example, the weight of the product value feature is 0.2, the weight of the user preference feature is 0.25, the weight of the holiday attribute feature is 0.2, the weight of the market trend feature is 0.15, the weight of the supply chain status feature is 0.1, and the weight of the promotion strategy feature is 0.1. After normalization of each feature category, the fusion calculation is performed according to the weight.

[0128] Vector Generator 203: Maps the fused features into a six-dimensional feature vector, corresponding to product value, user preference, holiday attributes, market trends, supply chain status, and promotion strategy. For example, the six-dimensional feature vector of a product might be (0.92, 0.75, 0.85, 0.68, 0.9, 0.7).

[0129] 2. Holiday Graph Construction and Model Training (30 Days Before the Spring Festival)

[0130] 1. Working of Festival Map Generation Engine 3

[0131] Triple recognition component 301: An entity recognition algorithm is used to extract the product ID, holiday type (Spring Festival), and user ID entities from the six-dimensional feature vector. A relationship extraction model is then used to identify the associations between these three entities, such as "User H purchased product I during Spring Festival" and "Product J is suitable for Spring Festival." Furthermore, a graph neural network is used to calculate the product-holiday association, constructing a heterogeneous information network with the product's node attributes (value score, category), the holiday's node attributes (type, influence coefficient), and the user's node attributes (preference feature vector) as nodes, and "User purchased product" and "Product suitable for Spring Festival" as edges. The edge weight for "User purchased product" is calculated as purchase frequency divided by total purchases. For example, if user H purchased product I 5 times and 20 times during Spring Festival over the past three years, the edge weight is 5 / 20 = 0.25. The edge weight for "Product suitable for Spring Festival" is calculated as sales during Spring Festival divided by sales outside Spring Festival. For example, if product J sold 1,000 units during Spring Festival and an average of 200 units per month during non-holiday periods, the edge weight is 1,000 / 200 = 5. The network is then trained using a graph convolutional network. The node feature vector is input, and the node embedding representation is learned through two convolutional layers. The cosine similarity of the product-festival node pair is output as the association score. For example, the association score between product J and the Spring Festival is 0.85.

[0132] Graph storage component 302: Uses a graph database (e.g., Neo4j) to store ternary relationships and constructs a heterogeneous product-festival-user network using a node-edge structure. Products, festivals, and users are considered nodes, with their attributes stored in the node properties. Relationships are considered edges, with edge weights stored according to the above calculation results.

[0133] Dynamic Update Component 303: Set a 7-day window and repeat the triplet identification process for newly added data. Using an incremental update mechanism, the weights of edges in the graph are adjusted. For example, as the Spring Festival approaches, new user behavior and order data are continuously generated. The weights of the edges associated with "products adapted for the festival" are recalculated and adjusted to ensure the graph reflects the latest associations.

[0134] 2. Product-Holiday Association Model Training: Historical order data is extracted from the knowledge base to construct a heterogeneous information network. Meta-paths for "product-user-holiday" and "holiday-product-user" are defined. A random walk algorithm is used to sample paths within the network and extract product-holiday co-occurrence patterns. A graph attention mechanism is used to calculate node importance. For each node, features of neighboring nodes are weighted and aggregated using the attention coefficient. Model parameters are optimized through backpropagation to improve the model's accuracy in predicting product-holiday associations.

[0135] 3. Marketing strategy generation and execution (15 days before Spring Festival to Spring Festival itself)

[0136] 1. The operation of the strategy generator 401 in the intelligent decision-making module 4 is based on the node association weights of the festival graph, combining the collaborative filtering algorithm and the graph attention mechanism to generate a specific marketing strategy.

[0137] Product combination solutions: Based on the relationships between products, festivals, and users, personalized product combinations are generated for different user groups. For example, for users who want to have family gatherings, a "New Year's Eve dinner ingredient package (including meat, seafood, and vegetables) + alcoholic beverages" combination is generated based on their historical purchase history and the correlation between products in the festival map. For users who want to celebrate the New Year, a "gift box (such as a nut gift box or a health product gift box) + exquisite packaging service" combination is generated.

[0138] Price Adjustment Factor: We develop different price adjustment factors for different users based on their price sensitivity. For users with high price sensitivity (e.g., a value above 0.8), the price adjustment factor is set between 0.8-0.85 (i.e., a discount of 80-85%); for users with low price sensitivity (e.g., a value below 0.3), the price adjustment factor is set between 0.9-0.95. We also consider the price fluctuation range in market trends to ensure that price adjustments remain within a reasonable range.

[0139] Channel placement priority: Determine channel placement priority based on user activity on social media and channel preferences as reflected in past orders. For example, for users who frequently browse products on Douyin, prioritize relevant product bundles and promotional information on the Douyin platform. For users who typically shop within the app, display product bundle information in personalized recommendations on the app homepage and in push notifications.

[0140] 2. The strategy execution system synchronizes the generated strategy information such as product combination scheme, price adjustment coefficient, and channel investment priority to the marketing activity system of the platform and each sales channel through an interface. The marketing activity system performs product combination onboarding, price adjustment, and channel promotion according to the strategy.

[0141] IV. Strategy Feedback and Optimization (During and after the Spring Festival)

[0142] 1. The feedback optimizer 402 receives user response data after the strategy execution, including user click volume, order volume, purchase amount, etc. For example, the click volume of a certain product combination on the Douyin platform is 500 times, the conversion rate is 15% (75 orders), and the average order value is 300 yuan. According to the reward function R = 0.3 × click volume + 0.5 × conversion rate + 0.2 × average order value, the reward value of this strategy is R = 0.3 × 500 + 0.5 × 75 / 500 × 100 + 0.2 × 300 = 150 + 7.5 + 60 = 217.5.

[0143] The gradient descent algorithm is used to adjust the parameters of the strategy generator 401. For example, for product combinations with low conversion rates, their weights in recommendations are reduced; for channels with high reward values, their investment priority parameters are increased. At the same time, the optimization results are synchronized to the dynamic update component 303 of the holiday graph generation engine 3, which adjusts the edge weights in the graph according to these information. For example, for products in product combinations with high conversion rates, their edge weights with the Spring Festival and related users are increased.

[0144] 2. Continuous optimization iteration: every 1000 new user response data is received, the feedback optimizer will prioritize experience replay mechanism, sort the state space (product node features, holiday node features, user node features), action space (push product combination, price adjustment value, channel selection), reward value R, etc. according to |R|, and prioritize sampling high reward experience to update network parameters, continuously optimize strategy generator parameters, and improve marketing strategy effect.

[0145] Through the above detailed application in the Spring Festival promotion activities, the system fully plays the functions of multi-source data collection, dynamic feature modeling, holiday graph construction, and intelligent decision optimization, and realizes accurate and efficient marketing based on the six elements of order transaction, effectively improving the order transaction volume and user satisfaction during the Spring Festival.

[0146] The above description is only a preferred embodiment of the present application; however, the protection scope of the present application is not limited thereto. Any skilled person in the art can make equivalent substitutions or changes to the technical solutions and improvement concepts of the present application within the technical scope disclosed by the present application, which should be covered by the protection scope of the present application.

Claims

1. A marketing strategy optimization management system based on the six elements of order closing, characterized by: It includes a multi-source data acquisition module (1), a dynamic feature modeling unit (2), a festival map generation engine (3) and an intelligent decision module (4); The multi-source data acquisition module (1) comprises: Behavioral data collection subunit (101): By deploying embedded code on shopping platform pages and social media interactive interfaces, it can capture the user's browsing trajectory, search keywords, product collection and purchase operation behavior logs in real time; Order data interface subunit (102): connects with the enterprise order management system through a preset data interaction protocol and periodically extracts structured data of historical orders; Commodity data extraction subunit (103): establishes a connection with the database of the commodity management system and extracts commodity attribute information through a data synchronization tool; Festival data crawling subunit (104): uses a distributed crawler framework to crawl festival-related data from designated network channels, and has a built-in data cleaning component to filter out invalid information; The dynamic feature modeling unit (2) comprises: Original feature extractor (201): performs field analysis on the behavior log and order data output by the multi-source data acquisition module (1), and extracts six types of original features of product value and user preference; Feature fusion processor (202): normalizes and fuses the six types of original features through a preset weight matrix; Vector generator (203): maps the fused features into six-dimensional feature vectors, wherein the six-dimensional feature vectors correspond to commodity value, user preference, holiday attribute, market trend, supply chain status, and promotion strategy respectively; The festival map generation engine (3) includes: Triple recognition component (301): extracting the commodity ID, festival type, and user ID entities from the six-dimensional feature vector using an entity recognition algorithm, and identifying the association between the three using a relationship extraction model; Graph storage component (302): uses a graph database to store ternary associations and constructs a commodity-festival-user heterogeneous network with a node-edge structure; Dynamic update component (303): sets a time window, repeats the triplet recognition process for newly added data, and adjusts the weights of edges in the graph through an incremental update mechanism; The intelligent decision-making module (4) comprises: Strategy Generator (401): Based on the node association weights of the festival graph, combined with collaborative filtering algorithm and graph attention mechanism, it generates product combination plans, price adjustment coefficients and channel placement priorities; Feedback optimizer (402): receives user response data after the strategy is executed, adjusts the parameters of the strategy generator (401) through the gradient descent algorithm, and synchronizes the optimization results to the dynamic update component (303) of the festival map generation engine (3); Each subunit of the multi-source data acquisition module (1) pushes data to the original feature extractor (201) of the dynamic feature modeling unit (2) through the HTTP protocol, the vector generator (203) of the dynamic feature modeling unit (2) transmits the six-dimensional feature vector to the triple recognition component (301) of the festival map generation engine (3) through the RPC interface, the map storage component (302) of the festival map generation engine (3) provides map data to the strategy generator (401) of the intelligent decision module (4) through a data subscription mechanism, and the feedback optimizer (402) of the intelligent decision module (4) feeds back strategy effect data to the dynamic update component (303) of the festival map generation engine (3) in real time through WebSocket.

2. The marketing strategy optimization management system based on the six elements of order closing according to claim 1 is characterized by: The processing flow of the festival data crawling subunit (104) is as follows: Crawl statutory holiday dates, adjustment arrangements, and folk activities descriptions from news websites and government announcement pages; Extract data on traditional festival consumption habits from folk culture websites; Crawl holiday-related text from social media platforms and transmit it to the sentiment analysis subunit. The sentiment analysis subunit uses the pre-trained BERT model to perform sentiment classification. The classification output probabilities include positive, negative, and neutral probabilities. The sales data for the same period are extracted from the enterprise historical database and transmitted to the time series analysis subunit. The time series analysis subunit uses ARIMA and SARIMA models to fit the data and output trend prediction parameters. The trend prediction parameters include mean, variance and periodic fluctuation coefficient.

3. The marketing strategy optimization management system based on the six elements of order closing according to claim 1 is characterized by: The original feature extractor (201) extracts six types of features by the following technical means: Product value characteristics: Calculate three core indicators: gross profit margin = (sales price - cost price) / sales price, inventory turnover rate = sales volume in the past 30 days / average inventory, and sales growth rate = (this month's sales volume - last month's sales volume) / last month's sales volume. A comprehensive score is obtained through weighted summation. User preference characteristics: Based on user behavior data and historical order data, a collaborative filtering algorithm is used to generate user preferences for product categories. The collaborative filtering algorithm is based on a user-product rating matrix, and cosine distance is used for similarity calculation. The deep learning model outputs user price sensitivity and brand loyalty. The user price sensitivity ranges from 0 to 1, with higher values ​​indicating greater price sensitivity. Brand loyalty is calculated based on the frequency of repeated purchases of the same brand. Holiday attribute characteristics: A classification algorithm is used to categorize holidays into "legal holidays," "folk holidays," and "seasonal holidays." The impact coefficient of holidays on sales of each product category is calculated based on historical sales data. Market trend characteristics: Connect to the industry database API to obtain competitor prices and promotional activity data. Combined with macroeconomic indicators, use the LSTM model to predict the market demand growth rate and price fluctuation range for the next 30 days. Supply chain status characteristics: Through real-time interfaces with inventory management systems and logistics systems, supply stability scores, replenishment cycles, and logistics delivery timeliness are obtained; Promotion strategy features: Extract historical promotion methods from the company's marketing activity database, compare the conversion rate improvement and average order value change rate of different methods through A / B testing, and generate applicable product categories and optimal discount ranges for each promotion method.

4. The marketing strategy optimization management system based on the six elements of order closing according to claim 1 is characterized by: In the triple recognition component (301), the process of calculating the commodity-festival association based on the graph neural network is as follows: Construct a heterogeneous information network: Node attributes of products, festivals, and users are used as nodes, and "user purchases products" and "product adaptation to festivals" are used as edges. The node attributes of products include value ratings and categories, the node attributes of festivals include types and influence coefficients, and the node attributes of users include preference feature vectors. The edge weight calculation formula for "user purchases products" is: edge weight = purchase frequency / total consumption times. The edge weight calculation formula for "product adaptation to festivals" is: edge weight = sales during festivals / sales outside festivals. The network is trained using a graph convolutional network: the node feature vector is input, the node embedding representation is learned through two layers of convolutional layers, and the cosine similarity of the product-festival node pair is output as the association score.

5. The marketing strategy optimization management system based on the six elements of order closing according to claim 1 is characterized by: In the strategy generator (401), the user profile and festival feature matching model process integrating the attention mechanism is as follows: Sort user historical behavior data by timestamp, use LSTM network to embed time series, and obtain user behavior time series feature vector; Festival feature processing: Festival sentiment, influence coefficient, and product attribute features are combined into a comprehensive feature vector through feature splicing; Multi-head attention mechanism processing: Set 8 attention heads to learn the association weights of temporal features and comprehensive features in different subspaces, and generate a recommendation vector after weighted summation.

6. The marketing strategy optimization management system based on the six elements of order closing according to claim 1 is characterized by: In the feedback optimizer (402), the push strategy dynamic optimization algorithm based on reinforcement learning is as follows: Define the state space: S = product node characteristics, holiday node characteristics, user node characteristics, and the action space: A = pushed product combinations, price adjustment values, and channel selection; Reward function: R = 0.3 × click volume + 0.5 × conversion rate + 0.2 × average order value; Adopting the Actor-Critic framework: the Actor network generates the action probability distribution, and the Critic network evaluates the state-action value; Prioritized experience replay mechanism: Sort the S, A, R, and S' experiences by |R|, prioritize sampling high-reward experiences to update network parameters, and iterate once every 1,000 new data points.

7. A marketing strategy optimization management method based on the six elements of order closing, comprising the marketing strategy optimization management system based on the six elements of order closing as described in any one of claims 1 to 6, characterized in that: The following steps are involved: S1. Build a knowledge base of six festival characteristics: S11: Each subunit of the multi-source data collection module collects data in parallel. After removing outliers through the data cleaning component, the data is stored in the data lake according to four categories: "user behavior," "historical orders," "product attributes," and "holiday characteristics." S12. The original feature extractor of the dynamic feature modeling unit reads data from the data lake and extracts six types of original features, which are standardized into a unified format; S13. Store the standardized features into the knowledge base according to the association key of "product ID-holiday ID-user ID"; S2. Training the product-festival association model: S21. Extract historical order data from the knowledge base and build a heterogeneous information network; S22. Define meta-paths: "product-user-festival" and "festival-product-user." Use a random walk algorithm to sample paths in the network and extract product-festival co-occurrence patterns. S23. Use graph attention mechanism to calculate node importance: for each node, aggregate neighbor node features weighted by attention coefficient, and optimize model parameters through back propagation; S3. Generate personalized recommendation list: S31. Sort the target user's historical behavior data by time and extract temporal features using an LSTM network. S32, extracting the emotional characteristics and influence coefficient of the current festival and the attribute characteristics of the candidate products, and integrating them into comprehensive characteristics; S33. Calculate the matching degree between user temporal features and product comprehensive features through a multi-head attention mechanism, and generate a top 30 recommendation list in descending order of matching degree. S4. Online optimization push strategy: S41. Push the recommendation list to the user contact channel and collect user response data in real time; S42, the feedback optimizer converts the response data into a reward value R, and the reinforcement learning algorithm updates the policy generator parameters; S43. Synchronize the optimized product-festival association weights to the festival graph generation engine, update the weights of the corresponding edges in the graph, and achieve closed-loop optimization of the strategy.

Images