A b2b liquor recommendation order generation method and system

Abstract

The application discloses a kind of B2B liquor recommendation order generation method and system.The method includes the following steps: step S1: multi-source data acquisition, construct multi-level merchant portrait;Step S2: time series collaborative hybrid model predicts SKU demand;Step S3: multi-constraint and multi-objective optimization, output the purchase recommendation order obtained by multi-objective optimization solution;Step S4: merchant feedback real-time update portrait and model;Step S5: double condition intelligent push trigger.The application improves the intelligent level, operation efficiency and business value of B2B liquor recommendation platform, effectively solves the technical problems in the prior art, such as insufficient merchant characterization dimension, low demand prediction accuracy, limited order generation capability, missing model update mechanism and mismatched push timing, and has the advantages of improving recommendation accuracy, automatically generating complete orders, dynamically optimizing models and intelligent pushing.

Description

Technical Field

[0001] This invention belongs to the field of intelligent recommendation technology in e-commerce, specifically relating to a B2B beverage recommendation order generation method and system. Background Technology

[0002] In the actual operation of B2B (Business-to-Business) e-commerce platforms for alcoholic beverages, merchant procurement recommendation technology has long faced multiple challenges.

[0003] Existing systems primarily rely on two types of recommendation mechanisms: First, rule-based recommendations, which rely solely on merchants' historical orders to encourage repeat purchases of the same SKUs or force the push of a uniform best-selling list across the platform. This results in all merchants, regardless of their business type, receiving identical recommendations, failing to reflect individual business differences. Second, single-product collaborative filtering recommendations, which rely solely on merchant similarity calculations for single-product recommendations, completely ignoring seasonal fluctuations, holiday sales cyclical changes, and temporal characteristics, leading to a lack of dynamic adaptability in the recommendation results. These traditional solutions can only push individual product information in a fragmented manner and cannot automatically generate structured, complete purchase recommendation orders. Merchants are forced to manually complete SKU (Stock Keeping Unit) selection, quantity configuration, and combination optimization, significantly increasing operational complexity and reducing procurement efficiency.

[0004] In the retail and e-commerce sectors, an SKU is the smallest unit used to uniquely identify a specific product. Different specifications, colors, packaging, flavors, etc., of the same product all correspond to different SKUs.

[0005] A thorough analysis of the existing technology reveals a serious deficiency in merchant profiling. The current system relies solely on historical transaction data to construct profiles, failing to effectively differentiate between various business types such as restaurants, KTVs, and convenience stores. Furthermore, it neglects crucial dimensions such as geographical location, seasonal climate influences, and price sensitivity stratification, resulting in highly homogenized recommended content that cannot meet the personalized purchasing needs of diverse merchants.

[0006] Secondly, there are systemic defects in the accuracy of demand forecasting. Because only a single algorithm model is used, it cannot coordinate the horizontal similarity of preferences among merchants with the vertical cyclical changes in the merchants' own sales, and it lacks a dynamic weight adjustment mechanism for new and old merchants. New merchants have significant prediction biases due to the sparseness of historical data, while old merchants are difficult to adapt to business changes due to the rigidity of the model.

[0007] Secondly, the order generation capability has fundamental limitations. The system fails to integrate multi-dimensional constraints such as real-time inventory status, platform promotion rules, and cold chain logistics fulfillment capabilities. It lacks a multi-objective optimization process, only outputting isolated single-item recommendations and failing to generate complete SKU combinations that can be directly ordered. Fourthly, the model update mechanism lacks a feedback loop. Merchant feedback on recommended orders, such as clicks, acceptances, and modifications to purchase quantities, is not effectively collected and utilized, resulting in the inability to dynamically correct merchant profiles and predictive models, leading to a long-term rigidity in the recommendation strategy.

[0008] Finally, the push strategy suffers from timing mismatch. Using a fixed-time batch push method fails to incorporate the unique operating cycle characteristics of each merchant (such as peak sales periods) and real-time inventory dynamics to trigger restocking reminders. This results in a disconnect between recommended content and merchants' actual purchasing needs, leading to consistently low conversion rates.

[0009] The aforementioned issues collectively restrict the level of intelligent services offered by B2B beverage platforms, and existing technologies urgently need improvement to address these problems. Summary of the Invention

[0010] In view of the above-mentioned deficiencies of the prior art, the first aspect of the present invention provides a method for generating B2B beverage recommendation orders, comprising the following steps: Step S1: Collect data from multiple sources and construct multi-level merchant profiles; After preprocessing the multi-source data, four types of features, including merchant type features, geographical location features, business cycle features, and historical transaction features, are extracted as the core dimensions of the merchant profile. The merchant profile is stored in a graph database, and a merchant similarity graph and a merchant-SKU association graph are constructed based on the merchant profile. Step S2: Predict SKU demand based on the merchant's new / old attributes combined with a time-series collaborative hybrid model; Based on the multi-level merchant profile in step S1, collaborative filtering and time-series prediction are performed; the new and old attributes of the merchants are determined according to the total historical order volume, and the fusion weight of collaborative filtering and time-series prediction is dynamically configured according to the new and old attributes. Collaborative filtering is emphasized for new merchants, and time-series prediction is emphasized for old merchants. After weighted fusion, the predicted demand of each SKU is output. Step S3: Multi-constraint and multi-objective optimization, output the recommended purchase order obtained after solving the multi-objective optimization; Based on the predicted demand for each SKU output in step S2, and combined with real-time inventory, promotional rules, and logistics fulfillment capabilities, a multi-dimensional constraint screening is performed to form a candidate set of SKUs available for sourcing. Using this candidate set as the optimization space, a multi-objective optimization algorithm is adopted, setting order amount and SKU coverage as gain indicators and delivery cost as loss indicators. The gain indicators are optimized to increase their values ​​during the solution process, while the loss indicators are optimized to decrease their values ​​during the solution process. The complete SKU combinations obtained through multi-objective optimization and their corresponding purchase quantities are output, forming recommended orders adopted by direct-supply merchants. Step S4: Merchant feedback updates profiles and models in real time; Collect feedback from merchants on four types of behaviors related to recommended orders: clicking, accepting, rejecting, and modifying purchase quantities, and make differentiated updates; synchronize the update parameters generated by various types of feedback to the merchant profiles and prediction models in the graph database to form an iterative closed loop of recommendation-feedback-update-re-recommendation; Step S5: Dual-condition intelligent push notification triggered; The system identifies peak sales periods for merchants and triggers periodic push notifications at a preset time before the peak sales period arrives. Simultaneously, it monitors merchants' own inventory data in real time and triggers inventory push notifications when the inventory falls below a preset safety threshold.

[0011] In the B2B beverage recommendation order generation method described above, optionally, in step S1, the preprocessing includes data cleaning and normalization; the multi-source data includes: collected merchant registration information, historical orders, map POIs, weather, and holiday multi-source data.

[0012] In the B2B beverage recommendation order generation method described above, optionally, in step S2, the collaborative filtering sub-model is used to mine the purchasing preferences of similar merchants and output a collaborative recommendation score; the time series prediction sub-model is used to split the trend component, seasonal component and holiday component to predict the sales volume of the target period.

[0013] In the B2B beverage recommendation order generation method described above, optionally, the collaborative filtering sub-model calculates cosine similarity based on merchant transaction features, and the time-series prediction sub-model adopts the Prophet algorithm.

[0014] In the B2B beverage recommendation order generation method described above, optionally, in step S3, SKUs are initially screened based on real-time inventory data to remove currently out-of-stock items; for the screened available SKUs, the platform's currently active discount and promotional activities are matched to calculate the discounted purchase price of each SKU; based on this, the fulfillment capabilities of the delivery area and cold chain logistics corresponding to each SKU are further evaluated to remove SKUs that cannot be fulfilled normally; using the SKU set after screening by triple constraints of inventory, price, and fulfillment as the optimization space, the NSGA-II genetic algorithm is used, with order amount and SKU coverage set as gain indicators and delivery cost set as loss indicator, outputting the complete SKU combination and its corresponding purchase quantity obtained through multi-objective optimization, forming a recommended order that can be directly adopted by merchants.

[0015] In the B2B beverage recommendation order generation method described above, optionally, in step S4, for the behavior of adopting a recommendation, the weight of the corresponding SKU and similar features in the merchant profile is increased; for the behavior of modifying the purchase quantity, the modified actual purchase quantity is used as the true value to reverse correct the parameters of the time series prediction sub-model; for the behavior of rejecting a recommendation, the rejected SKU and associated context are recorded and included in the negative sample set for periodic retraining of the collaborative filtering sub-model.

[0016] In the B2B beverage recommendation order generation method described above, optionally, in step S5, based on the business cycle characteristics included in the merchant profile constructed in step S1, the peak sales period of each merchant is identified, and a periodic push is triggered at a preset time point before the arrival of the sales peak, pushing the recommendation order generated in step S3 to the corresponding merchant; at the same time, the merchant's own inventory data is monitored in real time, and when the real-time inventory of any SKU is lower than the preset safety inventory threshold, a replenishment push is triggered immediately, pushing the replenishment recommendation order containing that SKU to the corresponding merchant; the periodic push and the inventory push complement each other, together constituting a dual-condition triggered intelligent push mechanism.

[0017] To achieve the above objectives, a second aspect of this application provides a B2B beverage recommendation order generation system, wherein the method for generating a B2B beverage recommendation order as described in any one of the first aspects comprises: The merchant profile building module is used to collect multi-source data, extract multi-dimensional features of merchants, build multi-level merchant profiles, and store them in the graph database. The demand forecasting module is connected to the merchant profile building module. It is used to perform collaborative filtering and time-series forecasting based on the merchant profile, and dynamically configure the fusion weight according to the new and old attributes of the merchants, and output the predicted demand of each SKU. The order generation module is connected to the demand forecasting module. It is used to perform multi-dimensional constraint screening based on the forecasted demand, combined with real-time inventory, promotion rules and logistics fulfillment capabilities, and solve the multi-objective optimization algorithm to generate a complete set of SKU combinations and purchase quantities obtained by multi-objective optimization, thus forming a recommended order. The push triggering module is connected to the order generation module and the merchant profile building module. It is used to obtain the merchant's operating cycle characteristics and real-time inventory data. When the preset cycle triggering conditions or inventory triggering conditions are met, the recommended order is pushed to the corresponding merchant. The feedback learning module, connected to the merchant profile building module, the demand prediction module, and the push triggering module, is used to collect merchants' behavioral feedback on recommended orders and update the model parameters of the merchant profile and the demand prediction module based on the behavioral feedback.

[0018] To achieve the above objectives, a third aspect of this application provides a terminal device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the B2B beverage recommendation order generation method as described in any one of the first aspects above.

[0019] To achieve the above objectives, a fourth aspect of this application provides a computer-readable storage medium storing computer-executable instructions or a computer program that, when processed and executed, implements the B2B beverage recommendation order generation method as described in any of the first aspects above.

[0020] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention constructs a four-layer multi-dimensional merchant profile and integrates multi-source data such as merchant registration information, historical orders, map POI, weather and holidays. It extracts four core dimensions: merchant type characteristics, geographical location characteristics, business cycle characteristics and historical transaction characteristics. It comprehensively depicts the business, location, cycle and transaction characteristics of merchants of different business formats. It fundamentally overcomes the shortcomings of traditional recommendation schemes, such as single merchant profile dimensions and serious homogenization of recommendation results, and greatly improves the accuracy of recommendation and the click intention of merchants.

[0021] (2) The present invention adopts a time-series collaborative hybrid prediction model, which integrates the horizontal purchasing preferences of similar merchants captured by the collaborative filtering sub-model with the vertical cyclical components of trends, seasons and holidays decomposed by the time-series prediction sub-model, and determines the new and old attributes of merchants based on the total historical order volume of merchants to dynamically adjust the weights of the two models, thereby achieving adaptive adaptation of data features of old and new merchants, effectively reducing the demand prediction error under the scenarios of seasonal transition and holiday fluctuations, and significantly improving the prediction accuracy compared with the single algorithm scheme.

[0022] (3) This invention integrates multiple fulfillment constraints such as real-time inventory, promotion rules and regional cold chain logistics delivery capabilities in the order generation process. It uses the SKU set after multi-dimensional constraint screening as the optimization space and adopts a multi-objective optimization algorithm to solve the problem with the goal of increasing order amount, optimizing SKU coverage and reducing delivery costs. It outputs a complete set purchase recommendation order containing SKU combination and purchase quantity, which replaces the traditional approach of only recommending individual products. While balancing platform revenue and fulfillment costs, it significantly increases the merchant's single purchase amount and the platform's overall GMV.

[0023] (4) This invention establishes a closed-loop self-learning mechanism for merchant behavior feedback. By collecting four types of behavioral feedback from merchants regarding recommended orders, namely, clicking, accepting, rejecting, and modifying the purchase quantity, the accepting behavior is used to increase the weight of the corresponding feature in the profile, the modifying behavior is used to correct the time series prediction parameters, and the rejecting behavior is included in the negative sample set for periodic retraining of the collaborative filtering model. The updated parameters are synchronized to the profile and the model, realizing automatic iterative optimization of profile weights and algorithm parameters. No manual intervention is required to adjust the recommendation rules, and the recommendation effect continues to improve as merchants use it.

[0024] (5) This invention innovatively proposes a dual-condition intelligent push mechanism. On the one hand, it identifies the peak sales period of each merchant based on the business cycle characteristics in the merchant profile and triggers a periodic push at a preset time before the peak to match the stocking rhythm. On the other hand, it monitors the merchant's own inventory in real time. When the inventory of any SKU is lower than the safety threshold, it immediately triggers a replenishment push to respond to the urgent procurement needs. It accurately captures the merchant's stocking peak period and inventory gap, effectively avoids the ineffective reach problem caused by traditional timed batch push, and significantly improves the merchant's user experience and order conversion efficiency.

[0025] In summary, this invention constitutes a complete technical loop from five stages: merchant profiling, demand forecasting, order generation, model iteration, and push notification. The synergistic effect of each stage systematically improves the intelligence level, operational efficiency, and commercial value of the B2B beverage recommendation platform. It effectively solves the technical problems in existing technologies, such as insufficient merchant profiling dimensions, low demand forecasting accuracy, limited order generation capabilities, lack of model update mechanisms, and mismatched push timing. It has the advantages of improving recommendation accuracy, automatically generating complete sets of orders, dynamically optimizing models, and intelligent push notifications. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the concept, specific structure and technical effects of the present invention will be further explained below in conjunction with the accompanying drawings, so as to fully understand the purpose, features and effects of the present invention.

[0027] Figure 1 This is a flowchart illustrating an embodiment of a B2B beverage recommendation order generation method according to the present invention; Figure 2 yes Figure 1 A detailed flowchart of one embodiment of step S1; Figure 3 yes Figure 1 A detailed flowchart of one embodiment of step S2; Figure 4 yes Figure 1 A detailed flowchart of one embodiment of step S3; Figure 5 This is a schematic diagram of the architecture of one embodiment of a B2B beverage recommendation order generation system of this application. Detailed Implementation

[0028] In this document, to make the technical means, inventive features, achieved objectives and effects of the invention readily understandable, the invention is further described below with reference to specific illustrations. However, the invention is not limited to the embodiments described below.

[0029] Terms such as “comprising” and “including” indicate that, in addition to the components that are directly and explicitly stated in the specification and claims, the technical solution of the present invention does not exclude the presence of other components that are not directly or explicitly stated.

[0030] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0031] In traditional B2B beverage platforms, merchant profiles are built solely based on historical transaction data, failing to integrate multi-dimensional features such as merchant type, geographical location, operating cycle, and historical transactions, resulting in a one-dimensional profile. Demand forecasting uses a single algorithm model, unable to dynamically adjust the weight allocation of collaborative filtering and time-series forecasting based on the merchant's newness or oldness, leading to low prediction accuracy. Recommendation systems only output discrete SKUs, without combining real-time inventory, promotional rules, and logistics fulfillment capabilities for multi-dimensional constraint filtering, failing to generate ready-to-order complete purchase recommendations. The model lacks a differentiated update mechanism for feedback on merchant clicks, acceptances, rejections, and changes in purchase quantities, resulting in a rigid recommendation strategy. Push strategies use fixed-time batch triggers, failing to correlate with merchant sales peaks and inventory status, leading to unreasonable push timing and impacting recommendation conversion rates. Specifically, the lack of profile dimensions prevents the system from distinguishing between different business types such as restaurants and KTVs; static prediction weights widen the demand gap between new and old merchants; the lack of constraints causes a disconnect between recommendation results and actual fulfillment capabilities; the lack of feedback mechanisms hinders model adaptive optimization; and the single push condition reduces the effectiveness of content reach.

[0032] For example, a hot pot restaurant located in a tourist area has historical order data showing a surge in demand for alcoholic beverages during holidays. However, the existing system only pushes generic SKUs based on the platform's best-selling list, failing to incorporate the seasonal fluctuations characteristic of the area. Furthermore, this restaurant is a new establishment, only six months old, with limited historical data. The predictive model over-relies on collaborative filtering, ignoring its rapid growth trend, resulting in product recommendations that don't match actual inventory needs. The restaurant must manually select SKUs and adjust quantities to generate orders, and the system continues to push recommendations at fixed times even when inventory is sufficient, leading to recommendations not being adopted. Further, in this scenario, the customer profile doesn't cover price sensitivity features, predictions lack dynamic weighting, constraints are missing, recommendations ignore logistics timeliness limitations, feedback behavior isn't collected and updated, and push notifications don't trigger inventory threshold monitoring, resulting in decreased conversion rates and increased manual workload.

[0033] If the above problems are not addressed, the recommendation system will be unable to provide accurate personalized recommendations, leading to a continuous decline in merchant satisfaction; the order generation process will rely on manual intervention, resulting in low operational efficiency and a high risk of errors; the model parameters will remain fixed for a long time, failing to adapt to the dynamic changes in merchant needs, leading to a degradation in recommendation quality; ineffective pushes will consume system resources and increase server load; ultimately, the platform's operational efficiency will decline, merchant stickiness will weaken, and the overall business sustainability will be affected.

[0034] In response, this application proposes a method for generating B2B beverage recommendation orders, such as... Figure 1 As shown, the specific steps may include the following: Step S1: Collect data from multiple sources and construct multi-level merchant profiles. A merchant profile is a data model that provides a multi-dimensional and structured description of a merchant. It integrates various features to form a multi-dimensional view of the merchant.

[0035] In such Figure 2 In the illustrated embodiment, after preprocessing the multi-source data, four types of features—merchant type characteristics, geographical location characteristics, operating cycle characteristics, and historical transaction characteristics—are extracted as the core dimensions of the merchant profile, which is then stored in a graph database. Based on this, a merchant similarity graph and a merchant-SKU association graph can be further constructed. For example, association rule analysis can be used to identify common purchasing patterns among merchants, or associations can be established by statistically analyzing the frequency of merchants' purchases of specific SKUs.

[0036] SKU is short for Inventory Unit, representing the smallest unit of a product in inventory that can be sold independently. In the beverage B2B scenario, SKU can refer to beverage products with specific specifications, brands, and packaging.

[0037] Graph databases are databases that use graph structures for data storage. Data is represented by nodes and edges; nodes represent entities (such as merchants or SKUs), and edges represent relationships between entities. Graph databases excel at handling relational queries and are suitable for building relational graphs.

[0038] A merchant similarity graph is a graph structure built based on various features in a merchant profile, calculating the similarity between merchants. In this graph, nodes represent merchants, edges represent the similarity relationships between merchants, and the weights of the edges indicate the degree of similarity.

[0039] A merchant-SKU relationship graph is a graph structure that describes the interaction between merchants and SKUs. In this graph, nodes can represent merchants and SKUs, and edges can represent merchant behaviors such as purchasing, browsing, or showing preferences for SKUs.

[0040] Step S2: Predict SKU demand using a time-series collaborative hybrid model.

[0041] In such Figure 3 In the embodiment shown, collaborative filtering and time-series prediction are performed based on the multi-level merchant profile in step S1; the new and old attributes of the merchants are determined according to their historical transaction data, such as the total number of historical orders, and the fusion weight of collaborative filtering and time-series prediction is dynamically configured. Collaborative filtering is emphasized for new merchants, and time-series prediction is emphasized for old merchants. After weighted fusion, the predicted demand of each SKU is output.

[0042] Collaborative filtering can recommend products based on merchants' shared purchasing behavior. For example, if both merchant A and merchant B purchased SKU X, and merchant A also purchased SKU Y, then SKU Y can be recommended to merchant B. Time-series forecasting can predict future sales based on historical sales data using moving averages or exponential smoothing. To improve forecast accuracy, new merchants and long-term partners can be differentiated based on their total historical order volume. For new merchants, due to their limited historical data, the results of collaborative filtering can be relied upon more heavily; for long-term partners, due to their abundant historical data, the results of time-series forecasting can be relied upon more heavily.

[0043] Step S3: Multi-constraint and multi-objective optimization, output the recommended purchase order obtained after solving the multi-objective optimization.

[0044] In such Figure 4 In the illustrated embodiment, based on the predicted demand for each SKU output in step S2, multi-dimensional constraints are applied to filter SKUs that are available for sale, taking into account real-time inventory, promotional rules, and logistics fulfillment capabilities. Using this SKU candidate set as the optimization space, a multi-objective optimization algorithm is employed, setting order amount and SKU coverage as gain indicators and delivery cost as loss indicators. The gain indicators are optimized to increase during the solution process, while the loss indicators are optimized to decrease during the solution process. The resulting complete SKU combinations and their corresponding purchase quantities, obtained through multi-objective optimization, form recommended orders directly adopted by merchants.

[0045] For example, inventory lists can be manually checked to remove SKUs that are currently out of stock; or SKUs participating in promotions can be manually selected based on platform-released promotional information; or SKUs that cannot be delivered on time can be excluded based on feedback from the logistics department. After these selections, a candidate set of available SKUs is formed. Using this candidate set as the optimization space, heuristic algorithms, such as greedy algorithms, can be used to solve the problem with the optimization objectives of increasing order value, optimizing SKU coverage, and reducing delivery costs. For example, SKUs with higher unit prices can be prioritized to increase order value, while trying SKUs containing diverse categories to increase coverage, and choosing warehouses closer to the merchant for delivery to reduce delivery costs. In this way, optimized SKU combinations and their corresponding purchase quantities are output, forming recommended orders that merchants can adopt.

[0046] Step S4: Merchant feedback updates profiles and models in real time.

[0047] The system collects feedback from merchants on four types of behaviors related to recommended orders: clicking, accepting, rejecting, and modifying purchase quantities, and updates these behaviors accordingly. The updated parameters generated by these feedbacks are then synchronized to the merchant profiles and prediction models in the graph database, forming an iterative closed loop of recommendation-feedback-update-re-recommendation.

[0048] Step S5: Triggering dual-condition intelligent push.

[0049] It identifies peak sales periods for merchants and triggers periodic push notifications at a preset time before the peak sales period arrives; at the same time, it monitors the merchant's own inventory data in real time and triggers inventory push notifications when the inventory falls below a preset safety threshold.

[0050] In this step, the system can manually identify peak sales periods for merchants based on their historical sales data, such as by observing sales curves over the past few months. At a preset time before the peak sales period, such as one day in advance, a periodic push notification is triggered, sending recommended orders to the merchant. Simultaneously, the system can monitor the merchant's own inventory data in real time, such as through inventory tables manually uploaded by the merchant. When it detects that the inventory of a merchant's SKU falls below a preset safety threshold, such as below 10 units, an inventory push notification is immediately triggered, sending a replenishment recommendation order containing that SKU to the merchant.

[0051] Furthermore, in step S1, preprocessing may include data cleaning and normalization, and multi-source data may include: collected merchant registration information, historical orders, map POIs, weather, and holiday multi-source data.

[0052] Merchant registration information describes the basic static attributes of merchants, such as their registered name, business type (e.g., restaurant, KTV, convenience store), registration date, and business address. This information forms the basis for building merchant profiles, distinguishing different types of merchants, and providing a macro-level classification basis for subsequent personalized recommendations. This data can be obtained directly from the B2B platform's merchant registration system or collected through interface integration with the business registration information database. Historical order data records merchants' past purchasing behavior, including the types, quantities, prices, purchase frequency, and order amounts of purchased SKUs. This data is crucial for analyzing merchant purchasing preferences, spending power, repurchase cycles, and identifying seasonal or holiday purchasing patterns. This data can be stored in the B2B platform's transaction database and periodically extracted and integrated through data interfaces or ETL (extraction, transformation, and loading) tools. Map POI (Point of Interest) data provides environmental information related to the merchant's geographical location, such as the type of business district (commercial area, residential area), distribution of surrounding competitors, transportation convenience, and pedestrian density. This information helps understand a merchant's operating environment, assess their potential customer traffic and sales capacity, and thus influence the recommendation strategy for beverage SKUs. This data can be obtained by calling the API interfaces of third-party map service providers (such as Gaode Maps and Baidu Maps) or through spatial data analysis and extraction using Geographic Information Systems (GIS). Weather data reflects the immediate impact of the external environment on beverage consumption, such as temperature, rainfall, and air humidity. Specific weather conditions may significantly affect the demand for certain beverage SKUs; for example, hot weather increases the demand for beer and iced drinks, while cold weather increases the demand for spirits and hot drinks. This data can be obtained through public data interfaces of meteorological bureaus or by subscribing to professional third-party weather data services. Holiday data includes information on national statutory holidays, traditional festivals, and local celebrations. Holidays can accompany consumption peaks or promotional activities for specific beverage SKUs, having a significant cyclical impact on merchants' purchasing needs. This data can be obtained from national statutory holiday databases, public calendar service APIs, or managed through manually maintained holiday lists.

[0053] Data cleaning aims to identify, correct, or remove inaccurate, incomplete, or irrelevant data from a dataset. Its purpose is to improve data quality and ensure the accuracy and reliability of subsequent analysis and model training. One approach is to detect outliers by defining a series of business rules and statistical thresholds. For example, for historical order data, upper and lower limits can be set for single purchase quantities or amounts; data exceeding these limits is marked as outliers and subject to manual review or automatic correction (such as replacement with the median or mean). Another approach is to utilize machine learning algorithms, such as Isolation Forest or Local Outlier Factor (LOF), to automatically identify outliers in the data and process them according to preset strategies, such as direct deletion, interpolation completion, or marking followed by manual confirmation. Normalization is the process of transforming data of different dimensions or numerical ranges to a uniform scale to eliminate the impact of differences in feature magnitudes on model training and ensure that all features have equal importance in the model. One approach is to use Min-Max normalization, linearly scaling the data to a fixed range (which could be [0, 1]), with the formula: X_norm = (X - X_min) / (X_max - X_min). Another approach is to use Z-score normalization (also known as standard deviation normalization), transforming the data into a distribution with a mean of 0 and a standard deviation of 1, with the formula: X_norm = (X - μ) / σ, where μ is the mean and σ is the standard deviation.

[0054] The following is a concrete example. Suppose a B2B beverage platform needs to generate recommended orders for a KTV located in a commercial area. First, the system collects the KTV's business registration information, such as its registration as an "entertainment venue," and obtains its detailed business address. Next, the system extracts the KTV's historical order data from the platform's database over the past year, including the types, quantities, purchase frequency, and total amount of SKUs (beer, spirits, soft drinks, etc.) purchased. Simultaneously, by calling a map service API, the system obtains POI information around the KTV, such as whether there are other KTVs, bars, or large shopping malls nearby, and the foot traffic level of its surrounding commercial area. Furthermore, the system also obtains historical weather data for the KTV's area, such as the average temperature and number of rainy days for each month over the past year, as well as the weather forecast for the coming week. Finally, the system integrates data on national statutory holidays and local festivals to identify dates that significantly impact KTV beverage demand. After this multi-source data collection is completed, the system performs preprocessing. For example, during the data cleaning phase, if any historical order data shows an abnormal purchase volume far exceeding the KTV's normal operating range (such as purchasing 1000 cases of beer at once), the system will mark it and verify or correct it to avoid erroneous data contaminating the merchant profile. During the normalization phase, data with different dimensions, such as order amounts (e.g., tens of thousands of yuan) and temperatures (e.g., 25 degrees Celsius), will be processed using Min-Max normalization to unify their value range to [0, 1]. This ensures that these features are treated fairly by the model when building the merchant profile, avoiding overemphasis or neglect of certain features due to differences in numerical values. In this way, the KTV's merchant profile will include information from multiple dimensions, such as its business type, geographical advantages, historical purchasing preferences, weather-related patterns, and holiday demand fluctuations, thus forming a comprehensive and accurate merchant profile.

[0055] Through the above technical solution, this application effectively solves the problems of single-dimensional merchant profiles and lack of depth and breadth in traditional solutions. By collecting multi-source data such as merchant registration information, historical orders, map POIs, weather, and holidays, and performing data cleaning and normalization preprocessing, the dimensions of the merchant profile are greatly enriched, enabling a comprehensive and accurate portrayal of merchants from multiple levels, including static attributes, dynamic behaviors, geographical environment, and external seasons. Data cleaning ensures the reliability of the profile data and avoids interference from outliers and noise on the accuracy of the profile; normalization eliminates the influence of differences in the units of measurement between different features, allowing the subsequent model to learn the influence of each feature on the merchant's purchasing needs more fairly and effectively. This high-quality, multi-dimensional merchant profile provides a more solid and accurate data foundation for the time-series collaborative hybrid model to predict SKU demand in the subsequent step S2, significantly improving the accuracy of demand prediction and the personalization of recommended orders. This enables the generation of recommended orders that better meet the actual needs of merchants, improving the adoption rate of recommendations and the operational efficiency of the platform.

[0056] In optional embodiments, such as Figure 3 As shown, in step S2, the collaborative filtering sub-model is used to mine the purchasing preferences of similar merchants and output a collaborative recommendation score; the time series prediction sub-model is used to split the trend component, seasonal component and holiday component to predict the sales volume of the target period.

[0057] This solution decomposes the SKU demand forecasting task into two complementary modules: a collaborative filtering sub-model and a time-series forecasting sub-model. These modules are then dynamically fused with merchant profiles to achieve more accurate demand forecasting.

[0058] The collaborative filtering sub-model aims to predict purchasing preferences by analyzing the similarity between merchants. This can be achieved by calculating the similarity between merchants based on their historical transaction records, browsing behavior, and geographical location, using methods such as cosine similarity or Pearson correlation coefficient. Based on this, a collaborative recommendation score is generated for the target merchant based on the purchasing behavior of other merchants similar to the target merchant. Another approach is to calculate the similarity between SKUs based on their shared purchasing patterns, and then recommend SKUs similar to those already purchased by the target merchant. The time-series prediction sub-model focuses on analyzing the time-series characteristics of merchants' historical sales data to predict future sales. This can be achieved by using an autoregressive moving average (ARIMA) model or a seasonal autoregressive moving average (SARIMA) model to predict future sales by identifying trends, seasonality, and cyclical patterns in historical sales data. Alternatively, exponential smoothing methods, such as the Holt-Winters model, can be used to capture trend and seasonal components by weighted averaging of historical data. This sub-model can break down complex sales fluctuations into trend components, seasonal components, and holiday components, thereby capturing the dynamic changes in sales more precisely.

[0059] The following is a concrete example. Suppose there is a newly opened bar on the platform with very limited historical transaction data. In this case, the collaborative filtering sub-model will identify other established bars that are highly similar to the new bar in terms of merchant type characteristics (e.g., both are bars) and geographical location characteristics (e.g., located in the same business district) based on the merchant profile built in step S1. By analyzing the historical transaction characteristics of these similar established bars, such as the beer brands and spirits they frequently purchase, the collaborative filtering sub-model can calculate and output a collaborative recommendation score for the new bar, such as recommending a craft beer and a whiskey. Meanwhile, for a chain convenience store that has been operating for many years, the time series prediction sub-model will conduct an in-depth analysis of its historical sales data. For example, for a popular carbonated beverage, the time series prediction sub-model will identify that its sales show a significant upward trend in the summer (trend component), sales on Fridays and Saturdays are significantly higher than Monday to Thursday (seasonal component), and sales peak during major holidays such as Spring Festival and National Day (holiday component). Based on these analyses, the time-series prediction sub-model can accurately predict the convenience store merchant's sales demand for the carbonated beverage in the coming week. In practical applications, the collaborative filtering sub-model can calculate cosine similarity based on merchant transaction characteristics to measure the similarity between merchants, while the time-series prediction sub-model can use the Prophet algorithm for prediction. Ultimately, considering the bar merchant's status as a "new merchant," the system assigns a higher weight to the collaborative filtering sub-model when fusing the prediction results from both models, ensuring that the recommended SKUs better match its potential purchasing preferences. For the convenience store merchant, being an "old merchant," the system assigns a higher weight to the time-series prediction sub-model, ensuring that the number of recommended SKUs more closely reflects its historical sales patterns and future predictions.

[0060] Through the above technical solutions, this application effectively solves the problems of traditional single prediction logic failing to fully cover the complex purchasing behavior of merchants and insufficient prediction accuracy. The collaborative filtering sub-model can effectively mine the purchasing preferences of similar merchants, especially in the case of new merchants or cold start scenarios, using collective intelligence to make up for the lack of individual data and provide accurate horizontal recommendation references. The time series prediction sub-model accurately captures the sales fluctuation patterns of merchants in different time dimensions by finely decomposing trend components, seasonal components, and holiday components, ensuring high accuracy in sales prediction for established merchants. This organic combination of collaborative filtering and time series prediction, and the dynamic configuration of fusion weights based on the new and old attributes of merchants, makes the prediction results reflect both the horizontal correlation preferences between merchants and the dynamic needs of merchants changing over time, thereby significantly improving the accuracy and robustness of SKU demand prediction. This provides a more reliable input for the multi-constraint and multi-objective optimization in the subsequent step S3, and thus can generate purchasing recommendation orders that are more in line with the actual needs of merchants and have greater commercial value after multi-objective optimization.

[0061] Furthermore, this application clarifies that the collaborative filtering sub-model calculates cosine similarity based on merchant transaction features, and the time-series prediction sub-model adopts the Prophet algorithm.

[0062] In this application, the core idea of ​​the collaborative filtering sub-model is to discover similarities between users or items based on user (in this case, merchants) behavioral data and make recommendations based on these similarities. It can be implemented in various ways, including user-based collaborative filtering, item-based collaborative filtering, and model-based collaborative filtering. This sub-model aims to identify groups of merchants with similar purchasing preferences or discover correlations between different SKUs by analyzing the purchasing behavior of a large number of merchants, thereby providing target merchants with SKU recommendations that they may be interested in. Merchant transaction characteristics refer to quantitative indicators extracted from merchants' historical transaction data that reflect their purchasing behavior patterns and preferences. These characteristics may include: the types, quantities, and frequencies of historically purchased SKUs, average order amount, repurchase rate of specific SKUs, SKU category distribution (e.g., the ratio of spirits, beer, and wine), purchasing time preferences (e.g., which days of the week and what times are most frequent for purchases), and interaction with promotional activities. These characteristics are the foundation for building merchant profiles and understanding merchant needs.

[0063] The Prophet algorithm works by decomposing time series data into trend terms, periodic terms (such as annual or weekly seasonality), and holiday effects. These terms can be modeled independently and then combined to form the final prediction. Its advantages include ease of configuration, robustness to missing and outliers, and the ability to intuitively interpret the various components of the prediction results.

[0064] Cosine similarity is a measure of the similarity between two non-zero vectors. It assesses their similarity by calculating the cosine of the angle between the two vectors. The closer the cosine value is to 1, the closer the directions of the two vectors are, meaning the more similar the features of the entities they represent (in this case, the merchants). In recommender systems, cosine similarity is often used to calculate the similarity between users or items because it is insensitive to the length of the vectors and focuses more on differences in direction, i.e., patterns of purchasing preferences rather than absolute purchasing quantities.

[0065] A time series forecasting sub-model is a model specifically designed for analyzing and forecasting time series data. Time series data is a collection of data points arranged in chronological order, such as historical sales data for SKUs. The goal of this sub-model is to identify patterns such as trends, seasonality, periodicity, and random fluctuations in the data, and to use these patterns to predict future values.

[0066] Through the above technical solution, this application effectively addresses the problems of relying solely on general algorithms to accurately capture the complex purchasing behavior characteristics of merchants and the lack of in-depth analysis of merchant transaction preferences and sales cycle patterns. Especially when facing sparse data for new merchants or complex sales fluctuations for established merchants, it can achieve more accurate SKU demand prediction. Ultimately, it provides more reliable and refined data support for subsequent multi-objective optimization to generate recommended purchase orders, significantly improving the adoption rate of recommended orders and merchant satisfaction.

[0067] In optional embodiments, such as Figure 4 As shown, in step S3, SKUs are initially screened based on real-time inventory data to remove currently out-of-stock items. Specifically, the system can query the warehouse management system (WMS) or supplier inventory interface in real time to obtain the current inventory level of each SKU, compare it with a preset minimum sellable inventory threshold, and mark SKUs with inventory levels below the threshold as out of stock and remove them. Alternatively, the platform can periodically synchronize inventory data from suppliers and store it in a local database, directly querying the local data during screening to exclude SKUs with zero or negative inventory.

[0068] Secondly, for the selected available SKUs, the system matches them with currently active discount and promotional activities on the platform to calculate the discounted purchase price for each SKU. This can be achieved by maintaining a promotional rule engine that includes rules for various promotional activities (such as discounts, bundled sales) and their effective times. The system applies all eligible promotional rules based on the SKU's attributes and the current time to calculate the final discounted purchase price. Alternatively, the platform can pre-configure all active promotional activities, their corresponding SKUs, and discount amounts. When calculating the purchase price, the system checks whether the SKU participates in the activity and calculates the discounted price based on the activity rules (e.g., a discount of Y yuan for purchases over X yuan or a direct Z-fold discount).

[0069] Based on this, further assessments are made of the delivery areas and cold chain logistics fulfillment capabilities corresponding to each SKU, filtering out SKUs that cannot be fulfilled normally. This can be achieved by leveraging the API of integrated logistics service providers, which allows real-time queries of the delivery range and fulfillment capabilities of logistics service providers based on the merchant's geographical location and SKU characteristics (such as whether cold chain is required, volume, and weight), removing SKUs that exceed the delivery range or cannot meet cold chain requirements. Alternatively, the platform can maintain a logistics fulfillment capability database, recording the service coverage and limitations for different delivery areas and different SKU types (such as ambient temperature, refrigerated, and frozen). The system can then determine whether an SKU can be fulfilled normally within the designated area based on the merchant's delivery address and SKU storage requirements.

[0070] After filtering based on inventory, price, and fulfillment constraints, the remaining set of SKUs is determined as the optimization space. This optimization space can be a list or dictionary containing SKUs and their discounted purchase price, delivery costs, etc., serving as input for a multi-objective optimization algorithm. Alternatively, a status field can be maintained for each SKU in the database, indicating whether it has passed the filtering process; optimization only uses SKU data with a status of "optimizable".

[0071] Within this optimization space, the NSGA-II genetic algorithm is used for solving the problem. The NSGA-II algorithm iteratively searches for a Pareto solution set through genetic operations such as generating an initial population, non-dominated sorting, crowding distance calculation, selection, crossover, and mutation. For example, existing open-source libraries (such as Python's DEAP library) can be used to configure the algorithm parameters and define fitness functions to evaluate the performance of the order scheme on each objective.

[0072] The optimization goals are set as increasing order value, optimizing SKU coverage, and reducing delivery costs. Increasing order value means increasing the total purchase quantity of all selected SKUs multiplied by their discounted purchase price. Optimizing SKU coverage can be defined as the proportion of selected SKUs to the total number of preferred SKUs in the merchant profile, or simply increasing the number of selected SKUs. Reducing delivery costs means reducing the total delivery cost of all selected SKUs. The NSGA-II algorithm will calculate these three objectives for each order scenario and find a balance between them, ensuring that no single objective is improved without compromising the others.

[0073] Ultimately, the algorithm outputs a complete set of SKU combinations and their corresponding purchase quantities obtained through multi-objective optimization, forming a recommended order that can be directly adopted by merchants. This can be achieved by selecting one or more order schemes from the Pareto solution set that meet specific business rules (e.g., total amount within a certain range, number of SKU types within a reasonable range), and outputting the SKU list and corresponding purchase quantities as a recommended order.

[0074] The following example illustrates this. Suppose a restaurant, based on previous forecasts, predicts its demand for 10 bottles of a certain brand of liquor (SKU_A) and 20 cases of a certain imported beer (SKU_B) for the following week. When generating a recommended order, the system first checks the real-time inventory and finds that only 5 bottles of SKU_A remain, while SKU_B has ample stock. At this point, the recommended quantity of SKU_A will be limited to 5 bottles, and the remaining 5 bottles will be marked as needing replenishment or recommended as alternatives. Next, the system detects that SKU_B is running a "buy 10 cases, get 1 free" promotion and calculates its discounted purchase price. Simultaneously, the system assesses the restaurant's location, finding it in a remote mountainous area. SKU_B requires cold chain transportation, but the current logistics provider cannot offer cold chain services in the region. Therefore, SKU_B is marked as unfulfillable and removed from the recommended order. Ultimately, the optimization space is limited to SKU_A (5 bottles). The NSGA-II genetic algorithm will build upon this foundation, aiming to increase order value (i.e., recommend 5 bottles of SKU_A), optimize SKU coverage (to meet merchant needs as much as possible), and reduce delivery costs. Since there is only one SKU, the algorithm will directly output a recommended order containing 5 bottles of SKU_A. This order is the result of multi-objective optimization after being filtered by triple constraints of inventory, price, and fulfillment, and can be directly pushed to the merchant for adoption.

[0075] Through the above technical solutions, this application effectively solves the problems of unexecutable orders, inaccurate pricing, or excessively high delivery costs caused by relying solely on predicted demand and ignoring actual business constraints in the traditional recommended order generation process. By introducing real-time inventory screening before multi-objective optimization, it ensures that all products in the recommended order are available, avoiding order invalidity or a decline in merchant experience due to stockouts. By matching platform promotional activities and calculating discounted purchase prices, the recommended order prices become more competitive in the market, increasing merchants' willingness to adopt them. By assessing logistics fulfillment capabilities, especially considering the specific delivery requirements of the beverage industry, it ensures that the recommended order is deliverable at the logistics level, avoiding order anomalies caused by logistics limitations. The introduction of these multi-dimensional constraints makes the conversion process from predicted demand to the final recommended order more closely resemble actual business scenarios, greatly improving the accuracy, feasibility, and economy of recommended orders. Combined with the accurate SKU demand forecasting in the preceding steps, the solution of this application can transform highly accurate forecast results into a truly executable, profitable, and efficient procurement recommended order, thereby significantly improving the overall quality of B2B beverage recommended orders and the procurement efficiency of merchants.

[0076] In step S4, for the behavior of adopting recommendations, the weight of the corresponding SKU and similar features in the merchant profile is increased; for the behavior of modifying the purchase quantity, the modified actual purchase quantity is used as the true value to reverse the parameters of the time series prediction sub-model; for the behavior of rejecting recommendations, the rejected SKU and associated context are recorded and included in the negative sample set for periodic retraining of the collaborative filtering sub-model.

[0077] When a merchant accepts an order recommended by the system or one of the SKUs included in it, it indicates that the SKU or its category aligns with the merchant's purchasing preferences. To strengthen the system's understanding of merchant preferences, various methods can be used to increase relevant weights. For example, a fixed weight increment can be set, and each time a merchant accepts an SKU, the weight of that SKU in the merchant's profile, as well as the weights of its category, brand, and other similar characteristics, are accumulated by this increment. Alternatively, weights can be adjusted using a dynamic weighting algorithm (such as exponential decay weighting) based on factors such as the frequency of acceptance behavior and order amount, so that recent or high-value acceptance behaviors have a greater impact on weight increases. Furthermore, machine learning models can be used to predict the impact of acceptance behavior on future purchasing preferences based on the merchant's historical behavior patterns, and weights can be adjusted accordingly. The merchant's modification of the SKU purchase quantity in the recommended order provides valuable information about the actual demand for that SKU. To improve the accuracy of the time-series prediction sub-model, the merchant's modified actual purchase quantity can be used as the true label for supervised learning. Specifically, the system can calculate the error between the purchase volume predicted by the time-series forecasting sub-model and the actual modification volume by the merchant. This error is then used to adjust the internal parameters of the time-series forecasting sub-model via backpropagation algorithms (such as gradient descent), including coefficients related to trend components, seasonal components, or holiday components. Another approach is to use online learning or incremental learning methods, using the actual purchase volume after each modification as new training samples to update the model parameters in real time, enabling the model to quickly adapt to changes in merchant needs. A merchant's refusal of recommended orders or specific SKUs clearly indicates their disinterest in those SKUs. To prevent the system from repeatedly recommending products the merchant dislikes, this information needs to be effectively utilized. The system can record the rejected SKU identifier, the specific time of the refusal, the merchant's profile characteristics at the time of recommendation, and the recommendation scenario (such as whether it is a promotional recommendation), and other related contextual information, marking this data as negative samples. This negative sample data can be periodically collected for retraining the collaborative filtering sub-model. During retraining, negative samples can be regarded as merchants' "dislike" or "low rating" of these SKUs. For example, in a collaborative filtering model based on matrix factorization, the value in the user-item interaction matrix corresponding to the negative sample can be set to 0 or a very small negative number, thereby guiding the model to reduce the priority of these SKUs in subsequent recommendations.

[0078] The following example illustrates this. Suppose a B2B beverage platform recommends a batch of beverage orders to a restaurant merchant. As a specific implementation method, if the restaurant merchant accepts the two baijiu (Chinese liquor) products "Wuliangye 52% ABV" and "Jiannanchun 52% ABV" included in the recommended order, the system will immediately increase the weight of these two SKUs in the merchant's profile. Simultaneously, since both liquors belong to the "high-end baijiu" category, the system will also correspondingly increase the weight of the "high-end baijiu" characteristic in the merchant's profile to reflect the merchant's preference for high-end baijiu. For example, if the predicted purchase quantity of "Qingdao Beer" in the recommended order is 10 cases, but the merchant modifies it to 8 cases based on actual inventory and sales expectations, the system will treat these 8 cases as the actual purchase quantity of "Qingdao Beer" and use this actual value to reverse-correct the prediction parameters related to "Qingdao Beer" in the time-series forecasting sub-model. For example, it may adjust its seasonality or trend predictive factors to make the model's future predictions closer to the merchant's actual needs. For example, if a merchant rejects a "French Bordeaux wine" item in a recommended order, the system will record the rejected SKU, along with contextual information such as the merchant type (restaurant), location, and time of recommendation. This information will be compiled into a negative sample set. When the collaborative filtering sub-model is periodically retrained, this negative sample set will be used for training, enabling the model to learn the negative preferences of that merchant or similar merchants for products like "French Bordeaux wine," thereby avoiding or reducing the recommendation of such products in future recommendations.

[0079] Through the above technical solution, this application effectively solves the problem that traditional recommendation systems lack a differentiation mechanism when processing merchant feedback, resulting in the model's inability to accurately distinguish the merchant's true intentions, and consequently leading to rigid recommendation strategies and insufficient personalization. The solution proposed in this application constructs a more intelligent and adaptive recommendation system capable of deep learning and optimization based on diverse merchant feedback, thereby continuously improving the adoption rate of recommended orders and merchant satisfaction, forming an efficient and accurate iterative closed loop of recommendation-feedback-update-re-recommendation.

[0080] In an optional embodiment, in step S5, based on the business cycle characteristics contained in the merchant profile constructed in step S1, the peak sales period of each merchant is identified, and a periodic push is triggered at a preset time point before the peak sales period, pushing the recommended order generated in step S3 to the corresponding merchant; at the same time, the merchant's own inventory data is monitored in real time, and when the real-time inventory of any SKU is lower than the preset safety inventory threshold, a replenishment push is triggered immediately, pushing the replenishment recommended order containing that SKU to the corresponding merchant; the periodic push and the inventory push complement each other and together constitute a dual-condition triggered intelligent push mechanism.

[0081] The merchant profile, which includes operational cycle characteristics, refers to information extracted from historical operational data to reflect a merchant's sales patterns, customer traffic trends, and inventory habits across different time dimensions (e.g., by week, month, quarter, or specific holidays). Its purpose is to provide the system with cyclical insights into merchant operations, enabling the prediction of potential demand. Specifically, this can be achieved by analyzing historical order data and statistically analyzing SKU sales volume for different dates, weeks, or months, thus constructing a time series of the merchant's sales volume. Furthermore, external data, such as local holiday schedules, information on major events, and even weather forecasts, can be correlated with historical sales data to more comprehensively depict the merchant's operational cycle characteristics.

[0082] Identifying peak sales periods for each merchant aims to determine specific timeframes during which sales significantly increase or reach peak levels, based on the operational cycle characteristics within the merchant profile. Its purpose is to provide an accurate time window for periodic push notifications, ensuring that recommended orders are delivered before the peak demand period. Specifically, based on the sales volume time-series data already constructed in the merchant profile, time-series analysis methods, such as seasonality decomposition, trend analysis, or outlier detection, can be used to automatically identify periods where sales are significantly above average. Alternatively, industry expert experience or preset rules can be combined to directly set peak sales periods for specific types of merchants (e.g., restaurants on weekends, KTVs on holidays). The preset time point triggering periodic push notifications refers to pushing recommended orders to merchants a preset time interval (e.g., N days or N hours in advance) before the identified peak sales period arrives. Its purpose is to give merchants sufficient time to prepare inventory, avoid stockouts during peak periods due to last-minute purchases, and improve order conversion rates. One approach is for the system to automatically trigger the generation and push of recommended orders when it detects that a merchant's peak sales period is approaching, based on a preset lead time (e.g., 3 days or 24 hours). Another approach is to dynamically adjust this preset time based on factors such as the procurement cycle of different SKUs, logistics delivery time, and the merchant's replenishment habits to ensure that recommended orders are delivered to the merchant in a timely manner.

[0083] Real-time monitoring of merchants' own inventory data refers to the system continuously acquiring and tracking the inventory information of the goods held by the merchants themselves. Its purpose is to promptly detect inventory shortages, providing data to trigger immediate replenishment notifications, thereby preventing merchants from missing sales opportunities due to insufficient inventory. Specifically, this can be achieved by establishing a data interface with the merchant's inventory management system (e.g., ERP system, POS system) to realize real-time synchronization or scheduled retrieval of inventory data. Additionally, the platform's data reporting function allows merchants to proactively update their inventory data, which the system then processes and monitors in real time. The preset safety stock threshold refers to a minimum inventory level set for each SKU. When the real-time inventory falls below this threshold, replenishment is considered necessary. Its purpose is to serve as a condition for triggering replenishment notifications, aiming to address demand fluctuations or supply uncertainties and ensure the continuity and stability of the merchant's inventory.

[0084] In practical applications, safety stock thresholds can be scientifically calculated using statistical methods (e.g., based on the standard deviation of demand forecasting errors) based on various factors such as historical sales volume of SKUs, supplier delivery cycles, merchant operating characteristics, and expected service levels. Alternatively, they can be set by merchants or platform operators based on experience and can be dynamically adjusted and optimized according to actual operating conditions.

[0085] Instant replenishment push notifications refer to the system immediately generating and pushing a replenishment recommendation order containing any SKU when the real-time inventory of any merchant's SKU falls below its preset safety stock threshold. Its purpose is to quickly respond to merchants' sudden replenishment needs, ensuring that merchants can replenish their inventory in a timely manner and avoid sales losses due to stockouts. One implementation method is to configure a background service that continuously polls or monitors inventory data changes; once inventory falls below the threshold, it immediately initiates the recommendation order generation process and pushes the order. Another implementation method is for the system to set a short check cycle (e.g., hourly or half-day), batch-evaluating the inventory data of all merchants within each check cycle and triggering replenishment push notifications for merchants that meet the criteria.

[0086] Specifically, periodic push notifications primarily address merchants' regular inventory needs during specific time periods (such as holidays and weekends), offering predictive capabilities; while inventory push notifications mainly address merchants' inventory shortages caused by daily sales, providing real-time updates. These two push methods operate independently yet work in tandem to ensure that recommended orders reach merchants at the most appropriate time, thereby improving the effectiveness of recommendations and merchant satisfaction.

[0087] Through the above technical solution, this application effectively solves the problems of inaccurate push timing and difficulty in accurately covering the ever-changing business needs of merchants under the traditional single push model. It not only improves the accuracy and timeliness of recommendations but also greatly enhances merchants' satisfaction and reliance on the platform, thereby promoting transaction activity on the B2B beverage platform.

[0088] This invention's solution is applied to online B2B wholesale platforms for alcoholic beverages, and is compatible with various offline alcoholic beverage purchasing merchants such as restaurants, community convenience stores, KTVs, and bars. In the specific implementation process, merchant registration information, historical orders from the past 12 months, business district POIs, weather data, and holiday data are collected. After cleaning and processing, multi-dimensional features are extracted and stored in the Neo4j graph database to construct merchant profiles and similarity maps.

[0089] Merchant similarity is calculated using cosine similarity, and the Prophet algorithm is used to fit sales cycle patterns. The weights of the two models are dynamically adjusted based on the merchant's age (new or old) to accurately predict the procurement demand for each SKU. Combining the platform's real-time inventory, promotional activities, and regional cold chain logistics fulfillment capabilities, the NSGA-II genetic algorithm solves the multi-objective final solution, generating a complete set procurement recommendation order that can be placed with one click. The system collects merchant operation behavior in real time, iterates and optimizes profiles and model parameters daily, sets differentiated push timings for merchants of different business types, and triggers emergency replenishment pushes based on inventory levels.

[0090] To verify the technical effectiveness of this invention, the solution was deployed on a B2B online wholesale platform for alcoholic beverages, and a three-month A / B comparison test was conducted. Participating merchants were randomly divided into an experimental group and a control group. The experimental group used the recommended orders generated by the solution of this invention, while the control group used the traditional best-selling list recommendation scheme. The two groups of merchants were evenly distributed in terms of business type, merchant size, and historical purchase amount.

[0091] The test results are shown in Table 1 below: Table 1 The test results above show that the solution of the present invention has achieved significant improvements in all core indicators compared with the traditional best-selling list recommendation solution, proving the actual technical effect of the present invention in improving recommendation accuracy, order conversion rate, merchant purchase amount and merchant satisfaction.

[0092] To achieve the above objectives, the present invention also provides a B2B beverage recommendation order generation system for implementing the B2B beverage recommendation order generation method as described in any of the preceding claims, which may include: The merchant profile building module is used to collect multi-source data, extract multi-dimensional features of merchants, build multi-level merchant profiles, and store them in a graph database.

[0093] The demand forecasting module is connected to the merchant profile building module. It is used to perform collaborative filtering and time-series forecasting based on the merchant profile, and dynamically configure the fusion weight according to the new and old attributes of the merchants, and output the predicted demand volume of each SKU.

[0094] The order generation module, connected to the demand forecasting module, is used to perform multi-dimensional constraint screening based on predicted demand, combined with real-time inventory, promotional rules, and logistics fulfillment capabilities. It then uses a multi-objective optimization algorithm to solve the problem and generate a complete set of SKU combinations and purchase quantities, forming a recommended order.

[0095] The push trigger module, connected to the order generation module and the merchant profile building module, is used to obtain the merchant's operating cycle characteristics and real-time inventory data. When the preset cycle trigger conditions or inventory trigger conditions are met, the recommended order will be pushed to the corresponding merchant.

[0096] The feedback learning module connects with the merchant profile building module, demand prediction module, and push triggering module. It is used to collect merchants' behavioral feedback on recommended orders and update the model parameters of the merchant profile and demand prediction modules based on the behavioral feedback.

[0097] Furthermore, in such Figure 5 In the illustrated embodiment, the merchant profile construction module may include a data acquisition unit, a feature engineering unit, and a profile storage unit. The data acquisition unit is used to collect merchant registration information, historical orders, map POIs, weather data, and holiday data; the feature engineering unit is used to preprocess the collected data and extract four types of features, including merchant type features, geographical location features, business cycle features, and historical transaction features, as the core dimensions of the merchant profile; the profile storage unit is used to store the merchant profile in a graph database and construct a merchant similarity graph and a merchant-SKU association graph.

[0098] The demand forecasting module can include a collaborative filtering unit, a time-series forecasting unit, and a fusion decision unit. The collaborative filtering unit calculates the similarity between merchants based on their transaction characteristics, mines the purchasing preferences of similar merchants, and outputs a collaborative recommendation score. The time-series forecasting unit decomposes historical sales data into trend, seasonal, and holiday components to predict sales for a target period. The fusion decision unit determines a merchant's new / old status based on their total historical order volume and dynamically configures the fusion weights of the collaborative filtering unit and the time-series forecasting unit, outputting the predicted demand for each SKU after weighted fusion.

[0099] The order generation module may include an SKU filtering unit, a price calculation unit, a delivery evaluation unit, and a combination optimization unit. The SKU filtering unit removes currently out-of-stock SKUs based on predicted demand and real-time inventory data. The price calculation unit matches the platform's currently effective promotional rules to calculate the discounted purchase price for each SKU. The delivery evaluation unit assesses the logistics fulfillment capabilities of the delivery area corresponding to each SKU and filters out SKUs that cannot be fulfilled normally. The combination optimization unit uses the SKU set filtered through multiple dimensions of inventory, promotion, and logistics constraints as the optimization space. It employs a multi-objective optimization algorithm, setting order amount and SKU coverage as gain indicators and delivery cost as loss indicators. The gain indicators are optimized towards increasing values ​​during the solution process, while the loss indicators are optimized towards decreasing values. The output is the complete SKU combination and purchase quantity obtained through multi-objective optimization.

[0100] The feedback learning module can include a behavior collection unit, a feature update unit, and a model iteration unit. The behavior collection unit collects four types of merchant feedback behaviors related to recommended orders: clicks, acceptance, rejection, and modification of purchase quantity. The feature update unit increases the weight of the corresponding SKU and similar features in the merchant profile for the behavior of accepting recommendations, and uses the modified actual purchase quantity as the true value to correct the time-series prediction parameters for the behavior of modifying purchase quantity. The model iteration unit records the rejected SKU and associated context for the behavior of rejecting recommendations, includes it in the negative sample set for periodic retraining of the collaborative filtering model, and synchronizes the updated parameters to the merchant profile construction module and the demand prediction module.

[0101] In other optional embodiments, the system may also include an effect evaluation module, which is connected to the push triggering module and the feedback learning module respectively, for statistical evaluation of the click-through rate, conversion rate and order amount of the recommended orders after the push, and feeds back the evaluation results to the feedback learning module to help optimize the profile weight and prediction parameters.

[0102] Other specific implementation methods have been described in detail above and will not be repeated here.

[0103] To achieve the above objectives, the present invention also provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor runs the program, it can implement the steps of the B2B beverage recommendation order generation method as described in any of the foregoing embodiments.

[0104] Processors and memory can be configured separately or integrated together, for example, integrated into a system-on-chip (SOC) of the terminal device.

[0105] To achieve the above objectives, the present invention also provides a computer-readable storage medium storing computer-executable instructions or a computer program, which, when processed and executed, implement the B2B beverage recommendation order generation method described above.

[0106] The computer-readable storage medium is, for example, memory. Memory can be volatile or non-volatile, or it can include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DRRAM).

[0107] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more computer devices (which may be personal computers, servers, or network devices, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0108] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for generating B2B beverage recommendation orders, characterized in that, Includes the following steps: Step S1: Collect data from multiple sources and construct multi-level merchant profiles; After preprocessing the multi-source data, four types of features, including merchant type features, geographical location features, business cycle features, and historical transaction features, are extracted as the core dimensions of the merchant profile. The merchant profile is stored in a graph database, and a merchant similarity graph and a merchant-SKU association graph are constructed based on the merchant profile. Step S2: Predict SKU demand based on the merchant's new / old attributes combined with a time-series collaborative hybrid model; Based on the multi-level merchant profile in step S1, collaborative filtering and time-series prediction are performed; the new and old attributes of the merchants are determined according to the total historical order volume, and the fusion weight of collaborative filtering and time-series prediction is dynamically configured according to the new and old attributes. Collaborative filtering is emphasized for new merchants, and time-series prediction is emphasized for old merchants. After weighted fusion, the predicted demand of each SKU is output. Step S3: Multi-constraint and multi-objective optimization, output the recommended purchase order obtained after solving the multi-objective optimization; Based on the predicted demand for each SKU output in step S2, and combined with real-time inventory, promotional rules, and logistics fulfillment capabilities, a multi-dimensional constraint screening is performed to form a candidate set of SKUs available for sourcing. Using this candidate set as the optimization space, a multi-objective optimization algorithm is adopted, setting order amount and SKU coverage as gain indicators and delivery cost as loss indicators. The gain indicators are optimized to increase their values ​​during the solution process, while the loss indicators are optimized to decrease their values ​​during the solution process. The complete SKU combinations obtained through multi-objective optimization and their corresponding purchase quantities are output, forming recommended orders adopted by direct-supply merchants. Step S4: Merchant feedback updates profiles and models in real time; Collect feedback from merchants on four types of behaviors related to recommended orders: clicking, accepting, rejecting, and modifying purchase quantities, and make differentiated updates; synchronize the update parameters generated by various types of feedback to the merchant profiles and prediction models in the graph database to form an iterative closed loop of recommendation-feedback-update-re-recommendation; Step S5: Dual-condition intelligent push notification triggered; The system identifies peak sales periods for merchants and triggers periodic push notifications at a preset time before the peak sales period arrives. Simultaneously, it monitors merchants' own inventory data in real time and triggers inventory push notifications when the inventory falls below a preset safety threshold.

2. The B2B beverage recommendation order generation method according to claim 1, characterized in that, In step S1, the preprocessing includes data cleaning and normalization; the multi-source data includes: collected merchant registration information, historical orders, map POIs, weather, and holiday data.

3. The B2B beverage recommendation order generation method according to claim 2, characterized in that, In step S2, the collaborative filtering sub-model is used to mine the purchasing preferences of similar merchants and output a collaborative recommendation score; the time series prediction sub-model is used to split the trend component, seasonal component and holiday component to predict the sales volume of the target period.

4. The B2B beverage recommendation order generation method according to claim 3, characterized in that, The collaborative filtering sub-model calculates cosine similarity based on merchant transaction features, and the time-series prediction sub-model uses the Prophet algorithm.

5. The B2B beverage recommendation order generation method according to claim 1, characterized in that, In step S3, SKUs are initially screened based on real-time inventory data to remove currently out-of-stock items; for the screened available SKUs, the platform's currently active discount and promotional activities are matched to calculate the discounted purchase price for each SKU. Based on this, the fulfillment capabilities of the delivery area and cold chain logistics corresponding to each SKU are further evaluated, and SKUs that cannot be fulfilled normally are screened out. Using the set of SKUs that have been screened by triple constraints of inventory, price and fulfillment as the optimization space, the NSGA-II genetic algorithm is used to set order amount and SKU coverage as gain indicators and delivery cost as loss indicator. The algorithm outputs the complete SKU combination and its corresponding purchase quantity obtained by solving the multi-objective optimization, forming a recommended order that can be directly adopted by merchants.

6. The B2B beverage recommendation order generation method according to claim 1, characterized in that, In step S4, for the behavior of adopting recommendations, the weight of the corresponding SKU and similar features in the merchant profile is increased; for the behavior of modifying the purchase quantity, the modified actual purchase quantity is used as the true value to reverse the parameters of the time series prediction sub-model; for the behavior of rejecting recommendations, the rejected SKU and associated context are recorded and included in the negative sample set for periodic retraining of the collaborative filtering sub-model.

7. The B2B beverage recommendation order generation method according to claim 5, characterized in that, In step S5, based on the business cycle characteristics contained in the merchant profile constructed in step S1, the peak sales period of each merchant is identified. A periodic push is triggered at a preset time point before the peak sales period, and the recommended order generated in step S3 is pushed to the corresponding merchant. At the same time, the merchant's own inventory data is monitored in real time. When the real-time inventory of any SKU is lower than the preset safety inventory threshold, a replenishment push is triggered immediately, and the replenishment recommended order containing that SKU is pushed to the corresponding merchant. The periodic push and the inventory push complement each other and together constitute a dual-condition triggered intelligent push mechanism.

8. A B2B beverage recommendation order generation system, characterized in that, The method for generating B2B beverage recommendation orders as described in any one of claims 1 to 7 includes: The merchant profile building module is used to collect multi-source data, extract multi-dimensional features of merchants, build multi-level merchant profiles, and store them in the graph database. The demand forecasting module is connected to the merchant profile building module. It is used to perform collaborative filtering and time-series forecasting based on the merchant profile, and dynamically configure the fusion weight according to the new and old attributes of the merchants, and output the predicted demand of each SKU. The order generation module is connected to the demand forecasting module. It is used to perform multi-dimensional constraint screening based on the forecasted demand, combined with real-time inventory, promotion rules and logistics fulfillment capabilities, and solve the multi-objective optimization algorithm to generate a complete set of SKU combinations and purchase quantities obtained by multi-objective optimization, thus forming a recommended order. The push triggering module is connected to the order generation module and the merchant profile building module. It is used to obtain the merchant's operating cycle characteristics and real-time inventory data. When the preset cycle triggering conditions or inventory triggering conditions are met, the recommended order is pushed to the corresponding merchant. The feedback learning module, connected to the merchant profile building module, the demand prediction module, and the push triggering module, is used to collect merchants' behavioral feedback on recommended orders and update the model parameters of the merchant profile and the demand prediction module based on the behavioral feedback.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor runs the computer program, it implements the B2B beverage recommendation order generation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions or computer programs, which, when processed and executed, implement the B2B beverage recommendation order generation method as described in any one of claims 1 to 7.

Images