E-commerce live broadcast real-time item selection method, device and system, and storage medium

Abstract

The invention discloses an e-commerce live broadcast real-time item selection method, device and system, and a storage medium, and the method comprises the steps: analyzing the matching degree of current live broadcast audience features and a candidate item list through employing multi-modal data and a large language model, updating an item selection list in real time in a live broadcast process through the recommendation capability of an offline model, and improving the user experience. And the conversion rate, ROI and decision-making efficiency of the live broadcasting room are improved.

Description

Methods, devices, systems, and storage media for real-time product selection in e-commerce live streaming Technical Field

[0001] This invention belongs to the field of information processing technology, and specifically relates to a method, device, system, and storage medium for real-time product selection in e-commerce live streaming. Background Technology

[0002] The e-commerce live streaming industry is undergoing a profound transformation from a "traffic-only" approach to a "quality-oriented" one. Data from Douyin's e-commerce platform in 2024 shows that influencer live streaming's contribution has dropped to 30%, with top influencers with over a million followers accounting for only 9% of GMV, while smaller influencers contribute 21%. Meanwhile, store-based live streaming sales accounted for over 65% of total sales, with over 1,000 merchants achieving over 100 million yuan in sales through store-based live streaming. This "decentralization" trend is even more pronounced on the Taobao platform—during the 2024 Double 11 shopping festival, store-based live streaming sales increased by 36% year-on-year, with 49 store-based live streaming rooms achieving growth rates exceeding 100%. Simultaneously, consumers are shifting from "following the streamer" to "following quality," with 72% of consumers focusing more on the product's cost-effectiveness than the streamer's personal influence. Against this backdrop, automated product selection technology is becoming the core engine of the e-commerce live streaming industry.

[0003] Automated product selection technology needs to balance efficiency and effectiveness. On the one hand, it needs to quickly select products from inventory to ensure the continuity of the live stream; on the other hand, it also has corresponding requirements for the monetization effect of the selected products, pursuing the maximum return on investment. Current mainstream automated product selection solutions mainly include: automated rule engines based on historical product selection data, such as selecting products for the next day based on indicators like profit margin greater than 20% and return rate less than 5%; automated product selection based on machine learning, evaluating product performance based on the day's live stream results and selecting those with higher final scores; automated product selection based on user attributes, selecting products that match the characteristics of the live stream audience, such as age, gender, and demographic attributes (tags like "stay-at-home mom" and "anime / manga"); and alternatively, combining the above data with recommendation algorithms to generate product selection results for the next day, reducing the workload of product selection.

[0004] However, all of the above solutions face challenges related to data drift, demographic changes, and traffic fluctuations. First, the historical performance of selected products cannot fully reflect their quality, and products may experience delayed conversions, leading to a risk of data reversal. Second, most current product selection solutions operate on a T+1 model, which is particularly problematic for small and medium-sized livestreamers. This lack of timely response to changing traffic and user structures can result in user churn. Finally, product selection decisions are affected by computing performance, hindering horizontal scaling and potentially causing the loss of high-quality products, thus hindering the livestream's growth. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method, apparatus, system, and storage medium for real-time product selection in e-commerce live streaming.

[0006] To achieve the above objectives, the present invention provides the following solution: a real-time product selection method for e-commerce live streaming, comprising: step 1, acquiring candidate product feature data; step 2, acquiring live stream user feature data and real-time online feature data; step 3, using the candidate product feature data, live stream user feature data, and real-time online feature data as input text for a large language model request, and performing online inference through prompt word templates; step 4, configuring subsequent product selection plans in the live stream backend based on the recommendation results and recommendation reasons returned by the model.

[0007] As a preferred option, the candidate product characteristic data includes: the proportion of GMV and the proportion of live streaming time within the category, the unit cost of live streaming within the category, the unit customer acquisition cost within the category, the unit GMV of live streaming within the category, the unit profit of live streaming within the category, the ARPU and ROI within the category.

[0008] As a preferred approach, the system retrieves the barrage content from the backend every 10 minutes via the live stream API, along with the current live stream audience's age / gender / regional distribution and the top 5 interest tags of the current live stream audience, as user attribute and interaction data. Real-time online feature data includes: the cumulative GMV increase percentage, cumulative profit increase percentage, inventory digestion rate change, the average sales, average GMV, and average profit margin of each SKU in the past 60 minutes, the average traffic consumption, average traffic consumption standard deviation, average traffic, traffic standard deviation, and SKU identifier for each SKU.

[0009] The present invention also provides a real-time product selection device for e-commerce live streaming, comprising: a first processing module for acquiring candidate product feature data; a second processing module for acquiring live stream user feature data and real-time online feature data; a third processing module for using the candidate product feature data, live stream user feature data, and real-time online feature data as input text for a large language model request, and performing online inference through prompt word templates; and a fourth processing module for configuring a subsequent product selection plan in the live streaming backend based on the recommendation results and recommendation reasons returned by the model.

[0010] As a preferred option, the candidate product characteristic data includes: the proportion of GMV and the proportion of live streaming time within the category, the unit cost of live streaming within the category, the unit customer acquisition cost within the category, the unit GMV of live streaming within the category, the unit profit of live streaming within the category, the ARPU and ROI within the category.

[0011] As a preferred approach, the system retrieves the barrage content from the backend every 10 minutes via the live stream API, along with the current live stream audience's age / gender / regional distribution and the top 5 interest tags of the current live stream audience, as user attribute and interaction data. Real-time online feature data includes: the cumulative GMV increase percentage, cumulative profit increase percentage, inventory digestion rate change, the average sales, average GMV, and average profit margin of each SKU in the past 60 minutes, the average traffic consumption, average traffic consumption standard deviation, average traffic, traffic standard deviation, and SKU identifier for each SKU.

[0012] The present invention also provides a real-time product selection system for e-commerce live streaming, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a real-time product selection method for e-commerce live streaming when executed by the processor.

[0013] The present invention also provides a storage medium storing a computer program, which executes a real-time product selection method for e-commerce live streaming when running.

[0014] Compared with existing technologies, the beneficial effects of this invention are as follows: 1. This invention fully utilizes the attribute characteristics and comment data of current livestream users, combined with the historical data performance of various products in the candidate product list, to update audience information in real time, accurately locate the candidate product list, improve the efficiency and accuracy of product selection, and achieve automated calculation of the best category, best livestream duration, and estimated ROI. The candidate product list is a set of selectable products in the backend of e-commerce livestreaming, and its size is generally measured in SKUs. This invention is applicable to livestreaming e-commerce scenarios with more than 2000 SKUs and more than 5 primary product categories.

[0015] 2. The prompt word template provided by this invention can be applied to real-time product selection in live e-commerce under various vertical scenarios, and has the ability to be horizontally expanded. One system can be used simultaneously for multiple live rooms and multiple product categories.

[0016] 3. This invention uses an open-source large language model with single-machine inference, which does not require access to paid APIs. It can complete the system launch and access at low cost and is suitable for various live streaming organizations that do not have software development capabilities. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 is a flowchart of the real-time product selection method for e-commerce live streaming according to an embodiment of the present invention; Figure 2 is the interactive interface and output results of the large language model; Figure 3 is the year-on-year change of ROI in the live streaming room during the experiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Example 1, as shown in Figure 1, provides a real-time product selection method for e-commerce live streaming, comprising: an offline data preparation stage and an online inference stage. The offline data preparation stage includes step 1) normalization and import of the candidate product list, and the online inference stage includes step 2) extraction and feature construction of user interaction data, step 3) model inference, and step 4) formatted model output and application in the live streaming room. The offline data preparation stage uses attribute data (category, color, specifications, third-party ratings), prices (cost price including tax, average selling price from third parties, profit range), and historical live streaming performance information (sales volume, average selling price, average number of viewers in the live streaming room, live streaming duration, and traffic costs) of various products from the live streaming room backend. The online inference stage uses real-time data obtained from the live streaming room backend, including user attribute data (number of viewers in the live streaming room, percentage of new viewers, age distribution, gender distribution, regional distribution, and interest tag distribution), user interaction data (text of bullet comments), and real-time sales data (sales quantity, GMV, sales profit, and current inventory of products currently being streamed). This real-time data is a backend application service provided by most live streaming platforms to assist broadcasters in product selection and optimization. The historical live stream performance information will be statistically analyzed for each live stream, and the corresponding indicators will need to be aggregated in the future.

[0022] Furthermore, the real-time product selection method for e-commerce live streaming specifically includes: Step 1) Candidate product list normalization and import, specifically including the following steps: Step 1-1) Export the key attribute data of the products from the data backend and store them in categories as candidate results for subsequent prompt word input. The organized product data should include category, attribute, price, and historical live streaming performance information. The categories will vary depending on the theme of the live streaming room and can be defined differently according to the actual live streaming room application of the system. In particular, this invention will use a 3C digital product live streaming room as an example to illustrate the application steps of the system. Step 1-1) The categories are mainly divided into computer category (computer, tablet), communication category (mobile phone), photography category (camera, action camera, lens), wearable device category (electronic watch, electronic bracelet, smart glasses), and consumer electronics category (TV, game console, Bluetooth headset). It is only necessary to specify the major category, but detailed SKU-level related data is required, that is, the various attribute characteristics of a single product, such as Apple AirPods 3.

[0023] Steps 1-2) Import the data into the real-time automatic product selection system and simultaneously calculate the proportion characteristics of each SKU, including the GMV proportion and live streaming duration proportion within the category. The calculated characteristics include the unit live streaming cost, unit customer acquisition cost, unit live streaming GMV, unit live streaming profit, ARPU, and ROI within the category. These eight key indicators provide quantifiable information for the model. The calculated characteristics are calculated using the following formulas: Unit live streaming cost = Σ (traffic cost) / Σ (live streaming duration); Unit customer acquisition cost = Σ (traffic cost) / Σ (average number of viewers in the live streaming room); Unit live streaming GMV = Σ (sales volume * selling price) / Σ (live streaming duration); Unit live streaming profit = Σ (sales volume * (selling price - cost price including tax)) / Σ (live streaming duration); ARPU = Σ (sales volume * selling price) / Σ (average number of viewers in the live streaming room); ROI = Unit live streaming revenue / Unit live streaming cost. The above calculated characteristics are then standardized within the category. The data standardization is as follows: Standardized Feature = (Original Feature - Average Feature Within Category) / Standard Deviation of Feature Within Category. These eight key features will serve as key features in the model's prompts, providing quantitative information while reducing the data size and the number of tokens the model needs to analyze. The above steps organize the data into a two-dimensional tabular structured data set. Each SKU, in addition to its own attributes, corresponds to one of the eight key features. This data is defined as candidate product feature data.

[0024] Step 2), user interaction data extraction and feature construction, specifically includes the following steps: Step 2-1), user data extraction. Every 10 minutes, the system uses the live stream API to retrieve user attribute and interaction data from the backend, including the most recent 30 minutes' worth of bullet comments, the current live stream audience's age / gender / regional distribution, and the current live stream audience's top 5 interest tags. The live stream API is a backend data extraction interface provided by the live e-commerce platform. User interaction data is unstructured text data.

[0025] Step 2-2) Real-time Sales Data Extraction. Similar to Step 2-1), real-time sales data is obtained via API, primarily describing the progress of the live stream. Specific data includes the sales quantity, GMV, sales profit, and inventory of the products currently being streamed, as well as live stream traffic and actual traffic consumption. Real-time sales data is also extracted every 10 minutes to form a sales snapshot. This sales snapshot is time-series structured data, with a set of records every 10 minutes, corresponding to the data performance of each product SKU during that time interval in the live stream.

[0026] Steps 2-3) Real-time Sales Feature Construction. The time-series data obtained in Step 2-2 is typically large in scale, consuming significant tokens and slowing down computation during large language model interactions. Therefore, feature construction is necessary for the time-series data. Specifically, this includes: Step 2-3-1) Basic Statistical Feature Extraction: For each SKU, calculate five dimensions for each SKU in every 10-minute time window: cumulative sales, cumulative GMV, cumulative profit, inventory at the end of the window, and average actual transaction price. If an SKU does not appear in the corresponding window (i.e., no sales), retain the data from the previous window, ensuring that each SKU has corresponding data in every time window.

[0027] Step 2-3-2) Construction of Long Short-Term Memory Features. The original time-series data is large in scale, requiring further extraction of statistical features. Based on the complete window-based statistical features obtained in Step 2-3-1), short-term and long-term memory features for each SKU are constructed. The short-term memory features mainly include the month-on-month features of the past 10-minute window, including the cumulative GMV increment percentage, cumulative profit increment percentage, and inventory turnover rate change for each SKU, totaling three dimensions. All short-term memory features are measured using percentages. The long-term memory features mainly include the statistical features of the past 60-minute window, including the average sales volume, average GMV, and average profit margin for each SKU in the past 60 minutes, totaling three features. In addition, since all SKUs have the same traffic and traffic consumption for the same time window, four additional features are calculated for the entire live broadcast room in the past 60 minutes: average traffic consumption, average traffic consumption standard deviation, average traffic, and traffic standard deviation. These are also used as long-term memory features for subsequent processing. These 10 features, combined with the SKU identifier, constitute the real-time online feature data.

[0028] Through steps 2-1) to 2-3), the basic data of the large language model can be updated every 10 minutes, which is used to submit corresponding data and requests in the form of dialogue to obtain the optimal product selection results for the next 10 minutes.

[0029] Step 3) The model inference uses the candidate product feature data provided in Step 1) and the live stream user feature data and real-time online feature data provided in Step 2) as the input text for the large language model request. Specifically, this includes the following steps: Step 3-1) Feature Token Compression. Generally, directly inputting two-dimensional table data into the model will consume a large number of tokens, affecting the model's inference speed. Data processing is necessary to reduce the number of tokens used. Token compression specifically includes the following steps: Step 3-1-1) Removing meaningless fields from the real-time online features, including columns marking time windows; Step 3-1-2) Simplifying complex field names in the candidate product feature data and real-time online feature data into a1, b1, c1, d1 and a2, b2, c2, d2, etc., and subsequently marking their corresponding meanings in the prompt words; Step 3-1-3) Multiplying all data features containing decimal points by 100 to remove the tokens occupied by the decimal points, and marking them in the feature prompt words; Step 3-1-4) Using the tab character "\t" to split all data.

[0030] Steps 3-1-1) to 3-1-4) can effectively reduce the token usage of the two-dimensional table data input into the model, enabling the model to analyze as much SKU information as possible.

[0031] Step 3-2) Prompt Construction. Unlike traditional end-to-end deep learning models, large language models are generative models, and the quality of the final output is influenced by the input. A set of prompts for a live-streaming e-commerce real-time automatic product selection system based on a large language model serves as an application template for this specific scenario. The prompts are generated by using a computer program to concatenate the data tables calculated in steps 1) and 2) into natural language text that the large language model can understand.

[0032] Specifically, taking a digital 3C live streaming room as an example, the prompt template used in this invention is: You are a product selection expert in live e-commerce, and I am currently live streaming. My current live streaming category is 3C digital products. I need you to analyze my offline SKU data and the real-time SKU characteristic data of my live stream today, combined with my current user composition and comment information, to calculate the list of products with the highest estimated ROI in the next 10 minutes. My output needs to include the recommended ID, the reason for the recommendation, and the estimated ROI, and should be output in JSON format.

[0033] I will provide you with three aspects of data for your reference, but to help you understand my data, I will first provide you with a data dictionary. The data field names are [az][x], where x=1 represents the offline attributes and characteristics of the SKU, with the corresponding fields in the following order: SKUID, category, color, specifications, cost, unit live streaming cost within the category, unit customer acquisition cost within the category, unit live streaming GMV within the category, unit live streaming profit within the category, ARPU within the category, and ROI; x=2 represents the user distribution data of the current live stream, with the corresponding fields in the following order: age / gender / regional distribution of the current live stream audience, top 5 interest tags of the current live stream audience, and the latest 10 bullet comments; x=3 represents the real-time data performance of each SKU up to the current point in time in this live stream, with the corresponding fields in the following order: SKU ID, month-on-month characteristics of the past 10 minutes—cumulative GMV increase percentage, cumulative profit increase percentage, inventory digestion rate change, average sales volume of the past 60 minutes, average GMV, average profit margin, average traffic consumption of the live stream in the past 60 minutes, standard deviation of average traffic consumption, average traffic, and standard deviation of traffic. The following is the specific data I am providing to you: [Table 1][Table 2][Table 3] Tables 1, 2, and 3, which correspond to the prompt word templates in step 3-1), require data processing according to step 3-1). The table data example in this embodiment only retains the first row; the actual list is consistent with the data compiled using steps 1) and 2) each time product selection is performed.

[0034] Table 1 Example – SKU Offline Attributes and Feature Data: a1\tb1\tc1\td1\tte1\tf1\tg1\th1\ti1\tj1\tk115233\tDigital 3C\tR\tBluetooth, Wireless, 4 Size\t152\t172\t1650\t7329\t1272\t5521\t112 Table 2 Example – Current Live Stream User Distribution Data: a2\tb2\tc2\td2\tte218~24:34&25~35:40\t40\tGuangdong:26&Jiangsu:12\tDigital, Camera, Mobile Phone, Smart Device, iPhone\tLowest Discount & Free Shipping & Warranty Period | Table 3 Example —Real-time data performance of each SKU: The key features in the prompts a3\tb3\tc3\td3\te3\tf3\tg3\th3\ti3\tj3\tk315889\t14\t21\t-15\t252\t682\t17\t1822\t2891\t5720\t2344 have all undergone token compression, therefore there are no decimal points, which can greatly reduce the number of tokens used. In particular, SKU IDs should be limited to the same major category, and the number should not exceed 300.

[0035] Step 3-3) Model Interaction. Deploy an open-source large language model on a local server and send keywords to the server via API. This invention is compatible with most mainstream open-source large language models, including but not limited to Qwen2.5, Qwen3, Deepseek-V3, and GPT-OSS. Specifically, Qwen3 is chosen as the base model for model interaction. Using the prompt word template from Step 3-2), fill in the corresponding data features, and the model can return recommended products and the reasons for the recommendations in JSON format, applying multimodal data features to the e-commerce real-time product selection system. Since feature extraction in Step 2) requires accumulating a certain amount of data, input to the model every 10 minutes during the live stream is sufficient. Depending on the token size and server performance, the response time is between 1 and 5 minutes. Specifically, the server used in this experiment is a single-machine Mac Studio, M3 Ultra 28 cores, 256 GB memory, which can support single-machine deployment and inference of the Qwen3 235B model. See Figure 2 for an example of the model interaction process.

[0036] Step 4) Formatting Model Output and Application in the Live Stream: Based on the recommendation results and reasons returned by the model, configure the subsequent product selection plan in the live stream backend. Specifically, the JSON data output by the model needs to be converted into a two-dimensional table-like data format so that operations personnel can quickly understand the reasons for product selection and the actions to be taken.

[0037] This invention fully utilizes product information, real-time sales data in the live stream, and user distribution information in the live stream backend. It can quickly update subsequent product selections by leveraging the multimodal data understanding and integration capabilities of a large language model, replacing manual or traditional algorithms. Furthermore, it can dynamically update the list based on the current user situation in the live stream. Compared to T+1 level product selection methods, it better balances flexibility and accuracy.

[0038] Taking the experimental results of this invention as an example, in the live streaming room of 3C digital products, the GMV of the live streaming room using this system increased by 8% year-on-year, and the ROI increased by 2% year-on-year. The comparison chart of the back-end data of the live streaming room is shown in Figure 3.

[0039] Example 2: This invention also provides a real-time product selection device for e-commerce live streaming, comprising: a first processing module for acquiring candidate product feature data; a second processing module for acquiring live stream user feature data and real-time online feature data; a third processing module for using the candidate product feature data, live stream user feature data, and real-time online feature data as input text for a large language model request, and performing online inference through prompt word templates; and a fourth processing module for configuring a subsequent product selection plan in the live streaming backend based on the recommendation results and recommendation reasons returned by the model.

[0040] As one embodiment of the present invention, the candidate product feature data includes: the proportion of GMV within the category and the proportion of live streaming duration within the category, the unit live streaming cost within the category, the unit customer acquisition cost within the category, the unit live streaming GMV within the category, the unit live streaming profit within the category, the ARPU within the category, and the ROI.

[0041] As one implementation of this invention, the system retrieves the barrage content of the most recent 30 minutes, the age / gender / regional distribution of the current live stream audience, and the top 5 interest tags of the current live stream audience every 10 minutes from the backend via the live stream API as user attribute and interaction data. Real-time online feature data includes: cumulative GMV increment percentage, cumulative profit increment percentage, inventory digestion rate change, average sales, average GMV, average profit margin of each SKU in the most recent 60 minutes, average traffic consumption, average traffic consumption standard deviation, average traffic, traffic standard deviation, and SKU identifier for each SKU.

[0042] Example 3: The present invention also provides a real-time product selection system for e-commerce live streaming, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a real-time product selection method for e-commerce live streaming when executed by the processor.

[0043] Example 4: The present invention also provides a storage medium on which a computer program is stored, wherein the computer program executes a real-time product selection method for e-commerce live streaming when it is running.

[0044] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for real-time product selection in e-commerce live streaming, characterized in that, include: Step 1: Obtain candidate product feature data; Step 2: Obtain user characteristic data and real-time online characteristic data for the live broadcast room; Step 3: Use candidate product feature data, live stream user feature data, and real-time online feature data as input text for the large language model request, and perform online inference through prompt word templates; Step 4: Based on the recommendation results and reasons returned by the model, configure the subsequent product selection plan in the live stream backend.

2. The real-time product selection method for e-commerce live streaming as described in claim 1, characterized in that, The candidate product feature data includes: the proportion of GMV and the proportion of live streaming time within the category, the unit cost of live streaming within the category, the unit customer acquisition cost within the category, the unit GMV of live streaming within the category, the unit profit of live streaming within the category, the ARPU and ROI within the category.

3. The real-time product selection method for e-commerce live streaming as described in claim 2, characterized in that, Every 10 minutes, the system retrieves the barrage content from the backend, the age / gender / regional distribution of the current live stream audience, and the top 5 interest tags of the current live stream audience as user attribute and interaction data via the live stream API. Real-time online feature data includes: cumulative GMV increment percentage, cumulative profit increment percentage, inventory digestion rate change, average sales volume, average GMV, average profit margin of each SKU in the last 60 minutes, average traffic consumption, average traffic consumption standard deviation, average traffic, traffic standard deviation, and SKU identifier.

4. A real-time product selection device for e-commerce live streaming, characterized in that, include: The first processing module is used to acquire candidate product feature data; The second processing module is used to acquire user characteristic data and real-time online characteristic data in the live broadcast room; The third processing module is used to use candidate product feature data, live stream user feature data, and real-time online feature data as input text for the large language model request, and to perform online inference through prompt word templates; The fourth processing module is used to configure subsequent product selection plans in the live streaming backend based on the recommendation results and reasons returned by the model.

5. The e-commerce live streaming real-time product selection device as described in claim 4, characterized in that, The candidate product feature data includes: the proportion of GMV and the proportion of live streaming time within the category, the unit cost of live streaming within the category, the unit customer acquisition cost within the category, the unit GMV of live streaming within the category, the unit profit of live streaming within the category, the ARPU and ROI within the category.

6. The e-commerce live streaming real-time product selection device as described in claim 5, characterized in that, Every 10 minutes, the system retrieves the barrage content from the backend, the age / gender / regional distribution of the current live stream audience, and the top 5 interest tags of the current live stream audience as user attribute and interaction data via the live stream API. Real-time online feature data includes: cumulative GMV increment percentage, cumulative profit increment percentage, inventory digestion rate change, average sales volume, average GMV, average profit margin of each SKU in the last 60 minutes, average traffic consumption, average traffic consumption standard deviation, average traffic, traffic standard deviation, and SKU identifier.

7. A real-time product selection system for e-commerce live streaming, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program performing the real-time product selection method for e-commerce live streaming as described in any one of claims 1-3 when executed by the processor.

8. A storage medium, characterized in that, The storage medium stores a computer program, which executes the real-time product selection method for e-commerce live streaming as described in any one of claims 1-3 when the computer program is running.

Images