Method for predicting on-shelf time of live commodities of e-commerce

Abstract

The invention discloses a method for predicting the on-shelf time of live commodities of e-commerce, and aims to improve the conversion efficiency and accuracy of live commodities. The method comprises the following steps: acquiring audience interaction data, including comments, likes, sharing times and watching duration, generated in a live broadcast process in real time through a preset data acquisition device or interface; the collected data are preprocessed, and audience behavior characteristics such as interaction frequency, participation depth and concern content are extracted; utilizing the constructed dynamic weight model to dynamically adjust the influence weight of different behavior characteristics on the shelf time decision, and combining real-time audience behavior characteristic data to generate a comprehensive time score for evaluating the commodity shelf potential at each time point; according to the comprehensive opportunity score, adjusting the commodity shelving opportunity in the live broadcast in real time, and recording conversion rate data after shelving; according to the invention, through combination of historical data and real-time behavior analysis, dynamic prediction of commodity on-shelf time is realized, and decision-making efficiency and commercial value of live broadcast e-commerce are significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of live broadcast technology, and in particular to a method for predicting the timing of listing products in an e-commerce live broadcast. Background Art

[0002] In recent years, with the rapid development of e-commerce live streaming, more and more businesses are promoting and selling products through live streaming platforms. In the existing technology, businesses usually rely on simple audience interaction data during live streaming, such as the number of comments and likes, and combine experience or static rules to determine the timing of product listing. These methods can improve the sales conversion rate of goods to a certain extent, but their accuracy and real-time performance are often subject to many limitations.

[0003] However, the existing technologies have the following major problems: insufficient depth in exploring audience behavior characteristics, and failure to effectively combine real-time data and historical data for dynamic analysis; lack of a dynamic adjustment mechanism for the weight of the impact of audience behavior characteristics on the timing of listing, resulting in the decision on the timing of listing relying on static rules or subjective judgment, which easily leads to missing the best sales opportunity; at the same time, conversion rate data is not effectively used to optimize subsequent decisions, and lacks self-improvement capabilities.

[0004] Therefore, there is an urgent need for a technical solution that can collect and analyze audience behavior characteristics in real time and dynamically predict the timing of product listing, so as to more accurately assist listing decisions in e-commerce live broadcasts. Summary of the invention

[0005] This application provides a method for predicting the timing of listing products in e-commerce live streaming, so as to improve the conversion efficiency and accuracy of live streaming sales.

[0006] This application provides a method for predicting the timing of listing of e-commerce live broadcast products, including: Acquire audience interaction data generated during the live broadcast in real time through a preset data collection device or interface, wherein the audience interaction data includes comments, likes, sharing times and viewing time sent by the audience; Preprocessing the collected audience interaction data to extract audience behavior characteristics, wherein the audience behavior characteristics include interaction frequency, participation depth, and attention content; Constructing a dynamic weight model for analyzing audience behavior characteristics, wherein the dynamic weight model is used to dynamically adjust the weight of the impact of different behavior characteristics on the decision of the time of listing; Using the dynamic weight model, combined with real-time audience behavior feature data, a comprehensive timing score is generated to evaluate the product listing potential at each time point during the live broadcast; According to the comprehensive timing score, the timing of product listing is adjusted in real time during the live broadcast, and the conversion rate data after listing is recorded to optimize the timing prediction in subsequent live broadcasts.

[0007] Furthermore, the e-commerce live broadcast product listing timing prediction method also includes dynamically adjusting the parameters in the weight model based on the conversion rate data and audience behavior characteristics to improve the timing prediction accuracy and conversion rate of future live broadcasts.

[0008] Furthermore, the real-time acquisition of audience interaction data generated during the live broadcast process through a preset data acquisition device or interface includes: Monitor the interactive behaviors in the live broadcast room in real time, including the frequency of comments, likes, shares, and bullet comments, and use them as initial behavioral data; Through the interface connection with the live broadcast platform, the monitored interactive data is securely transmitted to the data processing unit to ensure the real-time and integrity of data collection; The interactive data transmitted to the data processing unit is pre-processed, including deduplication, timestamp addition and data classification, to generate structured data.

[0009] Furthermore, the real-time acquisition of audience interaction data generated during the live broadcast process through a preset data acquisition device or interface includes: Capture audience behavior data in real time, including text interaction, emotional interaction and behavioral interaction; Label each interaction based on the type of data captured to differentiate the category and intent of the audience’s behavior; Stores tagged audience interaction data in a dynamic buffer.

[0010] Furthermore, the preprocessing of the collected audience interaction data to extract audience behavior characteristics includes: De-duplicate and denoise the collected audience interaction data, filter out invalid and abnormal data, and ensure data quality; The cleaned data is classified and processed according to interaction frequency, participation depth and attention content. The interaction frequency is calculated by counting the number of interactions of each viewer in unit time; the participation depth is determined by analyzing the audience's viewing time and the diversity of participation behavior; the attention content is clustered by extracting keywords from audience comments and barrages.

[0011] Furthermore, the construction of a dynamic weight model for analyzing audience behavior characteristics includes: The interaction frequency, participation depth and attention content in the audience behavior characteristics are classified and processed by the feature classification module of the dynamic weight model to generate a high-frequency feature vector, a depth feature vector and a content feature vector respectively; Using the feature weight allocation module of the dynamic weight model, the feature vector generated by the feature classification unit is preliminarily weighted to generate a preliminary weighted feature value, and the weight value of each feature is dynamically adjusted according to the correlation between the preset rules and the historical audience behavior data; Through the feature association analysis module of the dynamic weight model, the preliminary weighted feature value is associated with the historical behavior feature data, the behavior feature change trend in different time periods is analyzed, and the association feature matrix is ​​generated; Using the feature fusion module of the dynamic weight model, the high-frequency features, deep features and content features in the correlation feature matrix are fused to generate a comprehensive feature value, wherein the generation of the comprehensive feature value takes into account the dynamic impact of each type of feature on the conversion rate at the current moment; The comprehensive feature value is input into the timing scoring function through the scoring generation module of the dynamic weight model, and the comprehensive timing score is calculated in combination with the real-time data during the live broadcast, wherein the comprehensive timing score is used to evaluate the product listing potential at the current time point.

[0012] Furthermore, the feature weight allocation module using the dynamic weight model performs preliminary weighting processing on the feature vector generated by the feature classification unit to generate preliminary weighted feature values, including: According to the following formula 1, the high-frequency feature vector, the depth feature vector and the content feature vector generated by the feature classification unit are preliminarily weighted by a dynamic weighting function: ; in, Features The preliminary weighted eigenvalues ​​of ; Features The original eigenvalues ​​of Features The weight value is calculated according to the following formula 2: ; in, Features Real-time trigger frequency; , , , , , , Adjust parameters for presets; Features Relevance scores in historical audience behavior data; Features rate of change in real-time audience behavior data; Relevance score Calculate according to the following formula 3: ; in, Features At the point in time The audience behavior characteristic value; For time point Behavior weight factor; is the time span of historical data; Rate of change Calculate according to the following formula 4: ; in, Features The feature value at the previous time point; To avoid division by zero errors on positive numbers; Features The feature value at the current time point; is the magnification factor; Features The historical average characteristic value of is the adjustment factor; According to the following formula 5, generate the standardized weighted eigenvalue : ; in, is the total number of features.

[0013] Furthermore, the feature association analysis module of the dynamic weight model performs association calculation on the preliminary weighted feature value and the historical behavior feature data, analyzes the behavior feature change trend in different time periods, and generates an association feature matrix, including: Historical behavior feature data is analyzed at preset time intervals Split into multiple time periods , in Index for time period; According to the following formula 6, calculate the feature and In time period The strength of the association within : ; in, Features The preliminary weighted eigenvalues ​​of ; Features At the point in time The audience behavior characteristic value; is the time decay rate; For time period The central time point of Combine the correlation strength of all time periods to generate the final correlation feature matrix , whose elements Representation characteristics and Features The comprehensive correlation strength is calculated according to the following formula 7: ; in, Indicates the total number of time periods; Indicates time period The weight is calculated according to the following formula 8: ; in, Adjust parameters for weights; Indicates the index of the time period closest to the current time point of the live broadcast.

[0014] Furthermore, the feature fusion module using the dynamic weight model fuses the high-frequency features, deep features and content features in the associated feature matrix to generate a comprehensive feature value, including: The associated data of high-frequency features, deep features, and content features are extracted from the associated feature matrix and used as inputs for feature fusion, respectively, to ensure that each type of feature can independently reflect its role in audience behavior.

[0015] The weight of each type of feature is dynamically determined based on real-time audience behavior and historical behavior trends, ensuring that the feature categories that have a greater impact on conversion rate at the current moment have a higher weight in the fusion process.

[0016] The high-frequency features, deep features and content features are weightedly fused according to the assigned weights to generate comprehensive feature values.

[0017] Furthermore, the scoring generation module of the dynamic weight model inputs the comprehensive feature value into the timing scoring function, and calculates the comprehensive timing score in combination with the real-time data during the live broadcast, including: The comprehensive timing score is calculated according to Formula 9 below: ; in, is the number of dimensions of the comprehensive eigenvalue; The number of indicators for real-time data; Indicates Class comprehensive characteristic value; Indicates The weight of the class feature; Indicates Real-time indicator values; Indicates The weight of the real-time indicator; represents the weighted growth rate of audience behavior; and To adjust the parameters; Indicates the current live broadcast time point; The peak time point of behavior predicted by historical audience behavior data; To adjust the parameters.

[0018] The beneficial effects of the technical solution provided by this application include: The present invention obtains audience interaction data in real time during live broadcasts and analyzes it in combination with a dynamic weight model, so that the decision on the timing of product listing can respond to changes in audience behavior in real time, thereby improving the accuracy of timing prediction. The dynamic weight model adjusts the decision weight of the timing of listing according to the importance of the audience's behavioral characteristics, so that the effects of different characteristics in different live broadcast scenarios can be dynamically adapted, avoiding the limitations of traditional static rules. Dynamic adjustment and precise control avoid the energy waste caused by control delays and over-adjustment in traditional control methods, achieve efficient operation of the solenoid valve, help save energy and improve the overall efficiency of the system. 4 By recording the conversion rate data after listing and using it to optimize the model, the present invention has the ability of self-learning and improvement, can continuously improve the accuracy and applicability of product listing timing prediction, and provide reliable protection for long-term use. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flowchart of a method for predicting the timing of listing products in an e-commerce live broadcast provided in the first embodiment of the present application. DETAILED DESCRIPTION

[0020] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0021] The first embodiment of the present application provides a method for predicting the timing of listing products in e-commerce live broadcast. Figure 1 , which is a schematic diagram of the first embodiment of the present application. Figure 1 A method for predicting the timing of listing products in an e-commerce live broadcast is described in detail in the first embodiment of the present application.

[0022] Step S101: acquiring audience interaction data generated during the live broadcast in real time through a preset data acquisition device or interface, wherein the audience interaction data includes comments, likes, sharing times and viewing time sent by the audience.

[0023] In step S101, obtaining the audience interaction data generated during the live broadcast in real time through a preset data acquisition device or interface is an important basis for this embodiment. Specifically, firstly, a data acquisition device is integrated or connected to the live broadcast platform. The device can be a software interface (API) or a hardware sensor device, which is used to connect to the interactive data stream of the live broadcast platform in real time. The data acquisition device needs to communicate with the interactive system of the live broadcast platform in order to capture various interactive behaviors generated by the audience during the live broadcast in real time.

[0024] The interactive data obtained include but are not limited to the following categories: first, comments sent by viewers, which are usually in text form. The text information can be directly read through the data interface, and metadata such as timestamp and sender ID can be attached; second, like behavior, which is usually recorded as the click event of like and the time when it occurred; third, the number of shares, which involves the operation records of the audience sharing the live broadcast room within the live broadcast platform or on external social platforms. This part of data is obtained through the log function or event monitoring module of the live broadcast platform; fourth, viewing time, which is calculated based on the entry time and exit time of each viewer, and the cumulative viewing time of each viewer is updated in real time.

[0025] In order to ensure the real-time and completeness of data collection during implementation, a dynamic data buffer needs to be set up to receive the data stream transmitted from the collection device and refresh the latest data in the buffer at a predetermined time interval (such as every second). This dynamic buffer needs to support hierarchical storage of data, such as classified storage according to the type of interactive behavior, so that subsequent preprocessing steps can be efficiently called. In addition, the accuracy of the data needs to be considered during the data collection process, for example, through the deduplication mechanism to exclude duplicate data caused by network delays, or through the anomaly detection mechanism to filter abnormal audience behavior (such as machine-generated batch likes or comments).

[0026] For data transmission, it is recommended to use a secure transmission protocol (such as HTTPS or WebSocket) to ensure that the collected data will not be tampered with or lost during transmission. For scenarios with high real-time requirements, low-latency transmission mechanisms (such as event push based on message queues) can be preferred.

[0027] Through the above process, the real-time, accurate and complete audience interaction data is ensured, providing a reliable data foundation for behavioral feature extraction and dynamic weight modeling in subsequent steps.

[0028] Furthermore, the real-time acquisition of audience interaction data generated during the live broadcast process through a preset data acquisition device or interface includes: Monitor the interactive behaviors in the live broadcast room in real time, including the frequency of comments, likes, shares, and bullet comments, and use them as initial behavioral data; Through the interface connection with the live broadcast platform, the monitored interactive data is securely transmitted to the data processing unit to ensure the real-time and integrity of data collection; The interactive data transmitted to the data processing unit is pre-processed, including deduplication, timestamp addition and data classification, to generate structured data.

[0029] During the implementation process, it is first necessary to monitor the interactive behaviors in the live broadcast room in real time. This includes but is not limited to the frequency of audience comments, likes, shares, and bullet comments. By monitoring the user interaction events on the live broadcast platform, the time, frequency, and content of these behaviors can be captured. For example, each comment and bullet comment data needs to record the timestamp and sender ID at the same time, while the like and share operations need to count their cumulative number and time distribution. In order to ensure the integrity of the data, these initial behavior data need to be refreshed in real time to ensure that high-frequency interactions or sudden behaviors in a short period of time are not missed.

[0030] The monitored interactive data is connected and transmitted through the interface with the live broadcast platform. Data transmission needs to adopt a secure protocol (such as HTTPS or WebSocket) to prevent data from being tampered with or lost during transmission. The interface connection must also meet the requirements of low latency and high concurrency to adapt to the rapid changes in audience interactive behavior in live broadcast scenarios. During the transmission process, the system needs to verify the integrity of the data in real time, for example, by verifying the hash value of the data packet to confirm that the data has not been modified.

[0031] After being transmitted to the data processing unit, the interactive data needs to undergo a series of preprocessing operations to ensure that it can be used efficiently. The first step of preprocessing is deduplication. The system needs to delete duplicate interactive data by comparing timestamps and behavioral content, such as repeated comments or likes sent continuously by users. Subsequently, timestamps are added to accurately record the time of occurrence of each piece of data to provide support for subsequent time series analysis. Data classification requires different types of interactive behaviors to be divided into independent categories, such as comments, likes, and shares, which are stored in different classification buffers. In addition, barrage data needs to be further split into text content and related metadata (such as sender information).

[0032] The preprocessed data needs to generate structured data, which is usually stored in the form of tables or databases. These structured data must contain fields such as audience ID, behavior type, timestamp, and specific content of the behavior, so that the subsequent feature extraction module can access these data quickly and efficiently. Through the above process, the collected audience interaction data can meet the needs of e-commerce live broadcast scenarios in terms of real-time and completeness.

[0033] Furthermore, the real-time acquisition of audience interaction data generated during the live broadcast process through a preset data acquisition device or interface includes: Capture audience behavior data in real time, including text interaction, emotional interaction and behavioral interaction; Label each interaction based on the type of data captured to differentiate the category and intent of the audience’s behavior; Stores tagged audience interaction data in a dynamic buffer.

[0034] In actual implementation, it is first necessary to capture audience behavior data in real time. These behavior data are divided into three types: text interaction, emotional interaction, and behavioral interaction. Text interaction includes comments and barrage messages sent by viewers. The system captures each text content by monitoring the text input interface of the live broadcast platform, and records the sender ID and timestamp; emotional interaction is mainly reflected in the operation of likes, emoticons or other emotional expressions. These data are obtained by monitoring the click events or emotional feedback options in the live broadcast platform, and their occurrence frequency and time series distribution are counted; behavioral interaction includes the audience's sharing of live broadcast links, clicking on product links or entering shopping pages. The system captures these behaviors by tracking related trigger events and records the behavior targets (such as the specific products clicked or the channels shared).

[0035] The captured data needs to be further classified to provide clear behavioral categories and intentions for subsequent analysis. To this end, each audience interaction behavior will be labeled. The generation of labels is based on the type and characteristics of behavioral data. For example, the comments and barrages in text interactions can be identified through keyword analysis to identify their content intent (such as asking, expressing interest or suggestions), and add labels such as "inquiry", "interest" or "suggestion" to them; emotional interactions are directly classified according to the nature of the behavior, such as likes are labeled as "positive emotions", and expression operations can be further distinguished as "positive" or "negative" labels; behavioral interaction labels are generated according to the operation goals, such as clicking on product links will be labeled as "product interest", and sharing live broadcasts will be labeled as "dissemination behavior". These labels are automatically generated through matching rules and semantic analysis technology to ensure the consistency and accuracy of classification.

[0036] After the tags are added, all tagged audience interaction data will be stored in the dynamic buffer. The dynamic buffer is a real-time updated data storage module that can partition and store data according to data type and tag. To ensure real-time and efficiency, the buffer supports concurrent writing and fast query functions, allowing subsequent data analysis modules to access required data with minimal latency. The stored data not only contains the behavior type and tag, but also retains the content and time information of the original audience behavior, ensuring that the original data can be traced back for further analysis when necessary.

[0037] Through the above steps, the system can meet the actual needs of e-commerce live streaming in terms of real-time and data structuring, laying the foundation for subsequent behavioral feature extraction and analysis.

[0038] Step S102: pre-processing the collected audience interaction data to extract audience behavior characteristics, which include interaction frequency, participation depth and attention content.

[0039] In step S102, preprocessing the collected audience interaction data to extract audience behavior characteristics such as interaction frequency, participation depth, and content of interest is a key step in ensuring the accuracy of subsequent analysis. First, the audience interaction data collected in step S101 is stored in a structured data buffer to support data classification and hierarchical processing. At the beginning of preprocessing, the collected data should be quality checked, including removing duplicate data, filtering noise data, and repairing possible missing data. For example, for repeated like records, it is necessary to compare timestamps to determine whether they are the same operation; for garbled or meaningless characters in comment data, regular expressions can be used to clean them up.

[0040] After data cleaning, we start to classify the data. The extraction of interaction frequency can be achieved by counting the number of interactions of the audience in a unit time, such as counting the number of comments, likes and shares of each audience in one minute, and generating a time-related interaction frequency table. For the extraction of participation depth, it is necessary to analyze the behavior patterns of the audience, such as whether their viewing time is continuous, whether they participate in multiple interactive behaviors (such as commenting and liking at the same time), so as to generate a participation depth score. The score can be divided into multiple levels, such as low, medium, and high depth of participation, so as to assign weights in subsequent feature modeling.

[0041] To extract the content of interest, it is necessary to perform natural language processing (NLP) on the audience's text interaction data. First, the comments and bullet screen data are segmented, and after removing stop words, keyword matching or topic modeling technology (such as LDA model) is used to extract the core topics of the audience's attention. To improve accuracy, a domain dictionary can be established to associate the extracted keywords with the description of the product or the theme of the live broadcast to determine whether the audience's attention is highly related to the live broadcast content.

[0042] After completing the feature classification, the extracted features need to be standardized so that different features can be uniformly measured in subsequent analysis. During the standardization process, the interaction frequency and participation depth can be normalized and their values ​​can be adjusted to the range of 0 to 1; for the content of interest, it can be weighted according to the relevance score of the keyword match to obtain a comprehensive content of interest value. All processed feature data must be arranged in time series and stored in the feature buffer to support the construction of the subsequent dynamic weight model.

[0043] Through the above process, the audience's behavioral characteristics such as interaction frequency, participation depth and content attention are accurately extracted and quantified.

[0044] Furthermore, the preprocessing of the collected audience interaction data to extract audience behavior characteristics includes: De-duplicate and denoise the collected audience interaction data, filter out invalid and abnormal data, and ensure data quality; The cleaned data is classified and processed according to interaction frequency, participation depth and attention content. The interaction frequency is calculated by counting the number of interactions of each viewer in unit time; the participation depth is determined by analyzing the audience's viewing time and the diversity of participation behavior; the attention content is clustered by extracting keywords from audience comments and barrages.

[0045] First, the first step of preprocessing is to deduplicate and denoise the audience interaction data to ensure the quality of the data. Deduplication mainly targets repeated interactive behaviors, such as repeated comments or likes caused by network delays. These redundant data are identified and removed by comparing timestamps and behavior content. Denoising focuses on filtering invalid and abnormal data, such as empty comments without actual content, default prompt messages generated by the system, or abnormal user behaviors (such as batch likes or comments). In addition, non-language characters or garbled characters in barrages and comments can be identified and cleaned up through regular expressions to retain meaningful information.

[0046] Next, the cleaned data is classified according to the interaction frequency, participation depth and attention content. The classification of interaction frequency is achieved by counting the number of interactions of each viewer in a unit of time. For example, within a set time interval (such as one minute), the number of comments, likes and shares sent by the audience is counted. These frequency data can be used to measure the audience's activity level in the live broadcast room. The classification of participation depth is determined by analyzing the audience's viewing time and the diversity of participation behavior. Viewing time refers to the time that the audience stays in the live broadcast room, while behavior diversity refers to whether the audience performs multiple interactive operations at the same time, such as participating in comments, liking or sharing. Participation depth can be divided into three levels: high, medium and low, which are used to distinguish the degree of attention paid by different audience groups to the live broadcast content.

[0047] For the classification of the content of interest, it is necessary to conduct text analysis on the audience's comments and barrage content. First, the text is decomposed into independent words through word segmentation technology, and after removing stop words, keywords with high information content are extracted. These keywords are further grouped through the topic clustering algorithm to identify the core topics that the audience is concerned about. For example, when multiple comments mention the characteristics or discount information of a certain product, the topic can be marked as "product interest" or "promotion sensitive". In this way, the system can extract the audience's real interests from massive text data.

[0048] After the above steps, the audience behavior characteristics are converted into structured data, which can accurately reflect the audience's behavioral characteristics during the live broadcast. Each step of cleaning, classification and feature extraction is related to each other to ensure the accuracy and availability of the data processing results.

[0049] Step S103: constructing a dynamic weight model for analyzing audience behavior characteristics, wherein the dynamic weight model is used to dynamically adjust the weight of influence of different behavior characteristics on the decision of the time of listing.

[0050] In step S103, constructing a dynamic weight model for analyzing audience behavior characteristics is one of the core steps of the present invention. The model dynamically adjusts the influence weights of different behavior characteristics in decision-making to adapt to the real-time changes in audience behavior and the dynamic characteristics of the live broadcast scene.

[0051] First, the basic structure of the model needs to be initialized. The dynamic weight model includes multiple feature input nodes, each of which corresponds to an audience behavior feature extracted from step S102, such as interaction frequency, participation depth, and attention content. After each input node receives feature data, it is combined with the initial weight value associated with it. The weight value can be set according to the statistical laws in the historical data, for example, the interaction frequency is set to a high weight because it usually directly reflects the audience's willingness to interact.

[0052] The key to the model is the dynamic adjustment mechanism. Dynamic adjustment is achieved by real-time monitoring of the changing trends of audience behavior characteristics. For example, when the number of comments in the live broadcast room increases rapidly, the model automatically increases the weight of the interaction frequency node to enhance its influence in decision-making. In order to implement this dynamic adjustment mechanism, it is necessary to introduce weight update rules and calculate the importance changes of features in combination with real-time data. Weight updates can be based on the magnitude of changes in audience behavior, such as the growth rate or abnormal fluctuations of audience behavior characteristics.

[0053] In addition, the model needs to consider the correlation between features. By analyzing the correlation between different features, the dynamic weight model can assign the joint weight of interactive features. For example, when both the depth of engagement and the content of attention show a significant increase at the same time, the model can appropriately increase the combined weight of these two types of features to reflect the audience's higher attention to the live content.

[0054] In order to ensure the stability of the model, the dynamic weight model also needs to add a smoothing mechanism. The smoothing mechanism limits the speed of weight adjustment to avoid frequent changes in weights due to short-term fluctuations. In specific implementation, a time window can be introduced to perform sliding average on real-time changing data, thereby reducing the impact of burst data on model weights.

[0055] Finally, the dynamic weight model will output the weight distribution results, which will directly affect the subsequent calculation of the comprehensive timing score. By dynamically adjusting the weights, the model can respond to changes in audience behavior in the live broadcast scene in real time, providing accurate and real-time support for the decision-making of product listing timing.

[0056] Furthermore, the construction of a dynamic weight model for analyzing audience behavior characteristics includes: The interaction frequency, participation depth and attention content in the audience behavior characteristics are classified and processed by the feature classification module of the dynamic weight model to generate a high-frequency feature vector, a depth feature vector and a content feature vector respectively; Using the feature weight allocation module of the dynamic weight model, the feature vector generated by the feature classification unit is preliminarily weighted to generate a preliminary weighted feature value, and the weight value of each feature is dynamically adjusted according to the correlation between the preset rules and the historical audience behavior data; Through the feature association analysis module of the dynamic weight model, the preliminary weighted feature value is associated with the historical behavior feature data, the behavior feature change trend in different time periods is analyzed, and the association feature matrix is ​​generated; Using the feature fusion module of the dynamic weight model, the high-frequency features, deep features and content features in the correlation feature matrix are fused to generate a comprehensive feature value, wherein the generation of the comprehensive feature value takes into account the dynamic impact of each type of feature on the conversion rate at the current moment; The comprehensive feature value is input into the timing scoring function through the scoring generation module of the dynamic weight model, and the comprehensive timing score is calculated in combination with the real-time data during the live broadcast, wherein the comprehensive timing score is used to evaluate the product listing potential at the current time point.

[0057] The method for predicting the timing of product listing in e-commerce live streaming provided in this embodiment constructs a dynamic weight model for analyzing audience behavior characteristics, and uses a multi-stage processing method of classification, weighting, association and fusion to achieve dynamic analysis of audience behavior characteristics, and finally generates a comprehensive timing score for evaluating the potential for product listing.

[0058] First, the collected and preprocessed audience behavior feature data are classified and processed through the feature classification module of the dynamic weight model, and the interaction frequency, participation depth and attention content are converted into high-frequency feature vectors, deep feature vectors and content feature vectors respectively. The interaction frequency feature vector is based on the time series of statistical audience interaction frequency to capture the changing trend of behavior intensity; the deep feature vector is generated according to the audience's viewing time and behavior diversity in the live broadcast room, reflecting the audience's participation depth; the content feature vector is generated by analyzing the keywords in the comments and barrage, reflecting the audience's interests and attention topics. This step converts the original feature data into a structured multidimensional vector for subsequent processing.

[0059] Next, the feature weight assignment module of the dynamic weight model is used to perform preliminary weighting on the generated feature vectors. An initial weight value is assigned to each feature vector by calculating its correlation in the historical audience behavior data. The weight value assignment follows preset rules. For example, the weight of a high-frequency feature may be positively correlated with the magnitude of the change in the interaction frequency, while the weight of a deep feature may be affected by the distribution of viewing time. In real-time applications, weight assignment also needs to be adjusted dynamically. For example, when the growth rate of audience comments increases significantly, the weight value of the content feature can be increased to reflect its importance to the current decision. After preliminary weighting, preliminary weighted feature values ​​are generated for each feature, which provide the basis for subsequent association analysis.

[0060] The feature association analysis module of the dynamic weight model is used to associate the preliminary weighted feature values ​​with the historical behavior feature data to analyze the changing trends of the behavior features in different time periods. Specifically, the module compares the distribution differences between real-time data and historical data in the feature space to identify the changing patterns of audience behavior. For example, a significant increase in the correlation between high-frequency features and deep features in a certain period of time may indicate an increase in the audience's interactive interest in a particular product. The analysis results are output in the form of an association feature matrix, where each matrix element represents the comprehensive correlation strength between different features in a specific period of time.

[0061] Subsequently, the feature fusion module of the dynamic weight model is used to fuse the high-frequency features, deep features, and content features in the correlation feature matrix to generate a comprehensive feature value. The fusion process not only considers the current weight of each type of feature, but also analyzes the interaction between features. For example, when the correlation between high-frequency features and content features increases, the comprehensive feature value will give a higher weight to this correlation to more accurately reflect the audience's current behavior pattern. The comprehensive feature value is the final result of feature classification, weighting, and correlation analysis, and can fully describe the dynamic impact of audience behavior characteristics on conversion rate.

[0062] Finally, the comprehensive feature value is input into the timing scoring function through the scoring generation module of the dynamic weight model, and the comprehensive timing score is calculated in combination with the real-time data during the live broadcast. The scoring generation module will quantitatively evaluate the timing potential of the product listing based on the changes in the comprehensive feature value and real-time data. The generated comprehensive timing score is used to determine in real time whether the current time point is suitable for listing the product, and provide a basis for subsequent adjustments to the listing strategy.

[0063] Through this series of associations and hierarchical processing, the dynamic weight model achieves a comprehensive analysis and dynamic evaluation of audience behavior characteristics, ensuring the accuracy and adaptability of product listing timing predictions.

[0064] Furthermore, the feature weight allocation module using the dynamic weight model performs preliminary weighting processing on the feature vector generated by the feature classification unit to generate preliminary weighted feature values, including: According to the following formula (1), the high-frequency feature vector, depth feature vector and content feature vector generated by the feature classification unit are preliminarily weighted by a dynamic weighting function: ; in, Features The preliminary weighted eigenvalues ​​of ; Features The original eigenvalues ​​of Features The weight value is calculated according to the following formula (2): ; in, Features Real-time trigger frequency; , , , , , , Adjust parameters for presets; Features Relevance scores in historical audience behavior data; Features rate of change in real-time audience behavior data; Relevance score Calculate according to the following formula (3): ; in, Features At the point in time The audience behavior characteristic value; For time point Behavior weight factor; is the time span of historical data; Rate of change Calculate according to the following formula (4): ; in, Features The feature value at the previous time point; To avoid division by zero errors on positive numbers; Features The feature value at the current time point; is the magnification factor; Features The historical average characteristic value of is the adjustment factor; According to the following formula (5), the standardized weighted eigenvalues ​​are generated : ; in, is the total number of features.

[0065] For formula (1), the initial weighted eigenvalue middle, Features The preliminary weighted eigenvalue of represents the comprehensive influence of the feature at the current moment.

[0066] Features The original feature values ​​are directly provided by the audience behavior features extracted in step S102, specifically including interaction frequency, participation depth or attention content.

[0067] Features The dynamic weight value is calculated by formula (2).

[0068] Formula (2) describes Dynamic weight value calculation: ; in, Features The historical relevance score of is calculated by formula (3), which reflects the importance of the feature in the historical behavior data.

[0069] Features The real-time change rate of is calculated by formula (4) and is used to reflect the dynamic changes of this feature in the current audience behavior.

[0070] Features The real-time trigger frequency, that is, the cumulative number of times the feature is triggered per unit time, is directly obtained from the audience behavior data.

[0071] They are the weight coefficients of historical correlation, real-time change rate and the interaction between the two. The recommended values ​​are and 0.3, which can be fine-tuned according to the importance distribution of the actual features.

[0072] Used to control the strength and rate of the time decay function. Recommended values ​​are 0.1 and 0.05 to ensure that the time effect gradually weakens.

[0073] Used to adjust the periodic fluctuation influence of features. The recommended values ​​are 0.05 and 0.1, which can be flexibly adjusted according to the fluctuation period of features in the live broadcast scene.

[0074] In formula (3), Features At the point in time The audience behavior characteristic values ​​are directly obtained from historical data. For time point The behavioral weight factor is used to amplify the data weight of important time periods. It can be dynamically set according to the audience activity. The recommended value fluctuates between 0.5 and 1.5. For the time span of historical data, the recommended value is 7 days to capture the behavior change pattern within a week.

[0075] In formula (4), is the characteristic value at the current time point; is the characteristic value at the previous time point, obtained directly from the real-time data. Features The historical average characteristic value of . is the amplification factor, with a recommended value of 1.2, used to enhance the sensitivity of real-time changes. The recommended value for the adjustment factor is 0.8, ensuring that the dynamic weight changes are reasonable when the feature deviates too much from the historical mean. To avoid division by zero errors for small positive numbers, the recommended value is .

[0076] In formula (5), Features The initial weighted eigenvalues ​​of . is the total number of features, usually 3 (interaction frequency, engagement depth, and attention content).

[0077] Furthermore, the feature association analysis module of the dynamic weight model performs association calculation on the preliminary weighted feature value and the historical behavior feature data, analyzes the behavior feature change trend in different time periods, and generates an association feature matrix, including: Historical behavior feature data is analyzed at preset time intervals Split into multiple time periods , in Index for time period; According to the following formula (6), the feature is calculated and In time period The strength of the association within : ; in, Features The preliminary weighted eigenvalues ​​of ; Features At the point in time The audience behavior characteristic value; is the time decay rate; For time period The central time point of Combine the correlation strength of all time periods to generate the final correlation feature matrix , whose elements Representation characteristics and Features The comprehensive correlation strength is calculated according to the following formula (7): ; in, Indicates the total number of time periods; Indicates time period The weight is calculated according to the following formula (8): ; in, Adjust parameters for weights; Indicates the index of the time period closest to the current time point of the live broadcast.

[0078] First, the historical behavior feature data is collected at preset time intervals. Split into multiple time periods , each time period Includes a set of time points for aggregate analysis of feature data within a specific time interval. The time period should be dynamically set based on the actual live broadcast duration and the characteristics of audience behavior changes. For example, if the live broadcast duration is two hours and the audience interaction changes rapidly, It is set to 5 minutes so that each time period contains more detailed behavioral data.

[0079] In each time period Internally, calculate features and Features The strength of association , using the following formula: ; in, Representation characteristics The initial weighted eigenvalues ​​of are provided by the previous module, reflecting the characteristics The overall impact in the current time period.

[0080] Representation characteristics At the point in time The audience behavior characteristic values ​​are usually directly obtained from real-time or historical data to characterize the characteristics dynamic performance.

[0081] is the time decay rate, controlling the time point Relative to the center of the time period The recommended value is 0.05 to 0.1, with lower values ​​suitable for scenes with stable audience behavior and higher values ​​suitable for scenes with frequent behavior fluctuations.

[0082] It is the time period The central time point.

[0083] The denominator of formula (6) is obtained by and The square root of the sum is normalized to ensure the consistency of dimensions of different features.

[0084] Calculated The value represents the time period Internal, Features and Features The value range is usually between 0 and 1, which is used to quantify the dynamic relationship between features.

[0085] The feature correlation strength is integrated over all time periods to generate the final correlation feature matrix , whose elements Representation characteristics and Features The comprehensive correlation strength is calculated according to the following formula: ;

[0086] in, is the total number of time periods, which is related to the live broadcast duration and time interval Directly related.

[0087] It is the time period The weight of is calculated by formula (8), ensuring that the time period closer to the current time has a higher weight.

[0088] ;

[0089] It is the time period weight adjustment parameter, which controls the dynamic change rate of the time period weight. The recommended value is 0.1 to 0.5. A higher value is suitable for fast-changing live broadcast scenarios. Indicates the time period index to which the current time point belongs, which is usually calculated dynamically by the real-time system clock.

[0090] The model analyzes the dynamic correlation trends between features by integrating historical and real-time data. The hierarchical processing based on time periods ensures in-depth analysis of audience behavior in different periods, while the model's adaptability to time characteristics is enhanced through time decay and weight adjustment mechanisms.

[0091] Furthermore, the feature fusion module using the dynamic weight model fuses the high-frequency features, deep features and content features in the associated feature matrix to generate a comprehensive feature value, including: The associated data of high-frequency features, deep features, and content features are extracted from the associated feature matrix and used as inputs for feature fusion, respectively, to ensure that each type of feature can independently reflect its role in audience behavior.

[0092] The weight of each type of feature is dynamically determined based on real-time audience behavior and historical behavior trends, ensuring that the feature categories that have a greater impact on conversion rate at the current moment have a higher weight in the fusion process.

[0093] The high-frequency features, deep features and content features are weightedly fused according to the assigned weights to generate comprehensive feature values.

[0094] The method provided in this embodiment, through the feature fusion module of the dynamic weight model, fuses the high-frequency features, deep features and content features in the correlation feature matrix to generate comprehensive feature values, which is a key step in realizing the dynamic integration of audience behavior features. Specifically, it includes multiple stages of feature data extraction, weight allocation and feature weighted fusion to ensure that each type of feature accurately reflects its dynamic role during the fusion process.

[0095] First, the correlation data of high-frequency features, deep features, and content features are extracted from the correlation feature matrix. The correlation feature matrix is ​​generated by the feature correlation analysis module, and its elements represent the correlation strength of different features in different time periods. This step requires setting extraction rules for each type of feature. For example, high-frequency features are extracted through the audience's interactive behavior frequency data (such as likes, comments, and sharing times); deep features are obtained through the audience's viewing time and the depth of participation in various interactive behaviors; content features are obtained based on the keyword clustering results in the audience's comments or barrage. The extracted data are used as inputs to the fusion process, ensuring that each type of feature can independently reflect its specific role in the audience's behavior.

[0096] After the extraction is completed, the fusion weight of each type of feature needs to be dynamically determined based on the real-time audience behavior and historical behavior trends. The core of weight allocation is to comprehensively consider the immediate impact of real-time data and the long-term guiding role of historical trends. For example, at a certain point in time, if the number of likes and comments from viewers increases significantly, it indicates that high-frequency features have a greater impact on conversion rates. At this time, the weight of high-frequency features should be dynamically increased. If historical data shows that deep features (such as viewing time) contribute more to conversion rates at similar time points, the importance of the weight of deep features in fusion needs to be maintained. The dynamic allocation of weights can be achieved through preset weight adjustment rules. For example, the real-time change rate and historical correlation are added in a specific proportion to generate a dynamic weight for each type of feature.

[0097] Finally, the high-frequency features, deep features, and content features are weighted and fused according to the assigned weights to generate a comprehensive feature value. In the specific operation, the data of each type of feature is first multiplied by its corresponding dynamic weight to generate a weighted value, and then the weighted values ​​of all features are accumulated to obtain the final comprehensive feature value. The comprehensive feature value not only retains the independent role of different feature categories, but also reflects their relative importance at the current moment through dynamic weights, thereby providing accurate and comprehensive input for the subsequent comprehensive timing score.

[0098] Through the above-mentioned process of feature extraction, weight allocation and weighted fusion, the feature fusion module realizes the dynamic integration of audience behavior characteristics, so that the generated comprehensive feature value can fully reflect the audience behavior pattern at the current time point and its potential impact on the product conversion rate.

[0099] Furthermore, the scoring generation module of the dynamic weight model inputs the comprehensive feature value into the timing scoring function, and calculates the comprehensive timing score in combination with the real-time data during the live broadcast, including: The comprehensive timing score is calculated according to the following formula (9): ;

[0100] in, is the number of dimensions of the comprehensive eigenvalue; The number of indicators for real-time data; Indicates Class comprehensive characteristic value; Indicates The weight of the class feature; Indicates Real-time indicator values; Indicates The weight of the real-time indicator; represents the weighted growth rate of audience behavior; and To adjust the parameters; Indicates the current live broadcast time point; The peak time point of behavior predicted by historical audience behavior data; To adjust the parameters.

[0101] The prediction method provided in this embodiment uses the score generation module of the dynamic weight model to calculate the comprehensive timing score using formula (9). This formula combines comprehensive feature values, real-time data, and time dynamic factors to provide a comprehensive and accurate quantitative evaluation for the prediction of product listing timing. The following is a detailed description of each item in formula (9), including the calculation method and recommended value of the coefficient.

[0102] Formula (9) is: ;

[0103] In the molecular part, represents the result of weighting the comprehensive eigenvalues, is the dimension of the comprehensive feature value, usually including high-frequency features, depth features and content features, so .

[0104] It is A class of comprehensive feature values, generated by a feature fusion module, such as a weighted result representing a high-frequency feature. It is The weight of each class feature is used to measure the relative importance of each class feature to the score. Its calculation method can combine the change rate of real-time data and the relevance of historical data. The recommended value range is 0.3 to 0.5.

[0105] Represents the result of weighting the real-time indicator value. It is the number of indicators of real-time data, such as audience activity, comment frequency and click count, usually . It is A real-time indicator value comes directly from the real-time data collection module, such as the like frequency or the number of viewers at the current moment. It is The weight of each real-time indicator reflects the impact of each real-time data on the score. The recommended value is 0.2 to 0.6, and it is dynamically adjusted according to the changes in the real-time indicators.

[0106] In the molecule Part of the introduction is a growing trend in audience behavior.

[0107] The weighted growth rate of audience behavior is used to quantify the rate of change of audience behavior in the current time period. The calculation method is: ;

[0108] in, It's time The audience behavior value, is the time weighting factor. Indicates the total time span, usually the period covered by historical data. The value of can be calculated according to the audience activity, for example, a higher weight is assigned to the time period with a higher number of real-time online users.

[0109] is an adjustment parameter used to control the contribution strength of the growth trend, with a recommended value of 0.1 to 0.3.

[0110] is the amplification factor, used to enhance the growth rate The recommended value is 1.5 to 2.0.

[0111] The current live broadcast time point.

[0112] is a time adjustment parameter used to control the rate of time decay, with a recommended value of 0.05 to 0.1. Suitable for scenes with rapidly fluctuating behavior, lower values ​​for scenes with relatively stable behavior.

[0113] Finally, the calculation result of formula (9) is It is a comprehensive timing score that reflects the potential of the product to be put on the shelves at the current time. A high score indicates that the current time is suitable for putting the product on the shelves, while a low score indicates that the product will be put on the shelves later. This formula combines comprehensive feature values, real-time indicators and time dynamic factors to ensure that the scoring results are comprehensive and real-time.

[0114] Step S104: Utilize the dynamic weight model and combine it with the real-time audience behavior feature data to generate a comprehensive timing score for evaluating the product listing potential at each time point during the live broadcast process.

[0115] In step S104, the use of a dynamic weight model combined with real-time audience behavior feature data to generate a comprehensive timing score is the core step in realizing the potential assessment of product listing. First, the feature weight value at the current moment is obtained from the dynamic weight model constructed in step S103. These weight values ​​are the result of dynamic adjustment based on real-time audience behavior features, and are used to reflect the importance of each feature to the current live broadcast scene. At the same time, the audience behavior feature data extracted from step S102 also needs to be updated in real time to ensure the accuracy and real-time performance of the model input.

[0116] The generation of the comprehensive timing score is based on the combination of the dynamic weight model and real-time feature data. For each behavioral feature, such as interaction frequency, engagement depth, and attention content, it is necessary to associate and calculate the corresponding weight value. Specifically, after each type of feature data is weighted by the dynamic weight, the contribution value of this type of feature at the current moment is generated. The contribution values ​​of all features are further accumulated into a comprehensive quantitative result, which represents the overall audience behavior status at the current point in time.

[0117] At the same time, in order to ensure the accuracy of the scoring results, historical audience behavior data needs to be included as a reference factor in the scoring calculation. By analyzing the difference or trend change between the current feature data and the historical data, it is possible to evaluate whether the current audience behavior shows a positive interaction trend, such as a sudden increase in the frequency of interaction or a high level of attention to specific content. This information can be directly reflected in the score by adjusting the offset value of the dynamic weight.

[0118] The dynamic changes of live broadcast time must also be considered in the process of generating the score. In order to avoid large fluctuations in the score due to different time period selections, the score results can be smoothed by introducing a time decay mechanism. For example, as the live broadcast time increases, the impact of audience behavior characteristics on the score may weaken. By introducing time decay, the comprehensive timing score can more accurately reflect the product listing potential at the current moment.

[0119] The generated comprehensive timing score will be output in numerical form to quantify the product listing potential at the current time point. The scoring results can be used directly for decision-making and can also provide data support for subsequent timing optimization.

[0120] Step S105: According to the comprehensive timing score, the product listing timing is adjusted in real time during the live broadcast, and the conversion rate data after listing is recorded to optimize the timing prediction in subsequent live broadcasts.

[0121] In step S105, the timing of product listing is adjusted in real time according to the comprehensive timing score, and the conversion rate data after listing is recorded to further optimize the timing prediction model, which is an important part of the closed loop of the entire method. First, according to the comprehensive timing score generated in step S104, the potential for product listing at the current time point is judged. If the score exceeds the preset listing threshold, the product listing operation is triggered. After the product is listed, the system needs to immediately record the key conversion data after listing, including actual sales volume, click-through rate, audience comments and feedback, etc. These data will serve as the core basis for subsequent optimization.

[0122] In the process of adjusting the timing of product listing, it is necessary to ensure the real-time and accuracy of the operation. The system should continuously monitor the changing trend of the comprehensive timing score in the background, and dynamically decide whether to perform new product listing operations based on the real-time fluctuations of the score. For example, when the score shows a rapid upward trend but has not yet reached the threshold, the system can predict that the audience interaction enthusiasm may further increase, thereby postponing the listing time to obtain better conversion results. On the contrary, if the score has reached or exceeded the threshold and the audience interaction has stabilized, the product will be listed immediately to make full use of the current opportunity.

[0123] For the conversion rate data records after listing, the system needs to automatically collect relevant data and save it to the data storage module in time series. For example, record the sales volume and product click-through rate every minute to analyze the changes in audience behavior after listing. In addition, it is necessary to combine audience interaction data, such as whether the listing of products triggers more comments or discussions. These data can not only evaluate the listing effect of the current products, but also be used to improve the predictive ability of the comprehensive timing scoring model.

[0124] Through continuous monitoring and historical records of conversion rate data, the system can detect deviations between the score and the actual conversion effect. For example, if the score is high but the conversion rate is low, the prediction accuracy can be optimized by re-adjusting the influence weights of certain behavioral features in the weight model. Conversely, if the score is highly consistent with the conversion effect, the current model's ability to adapt to similar scenarios can be further enhanced.

[0125] Ultimately, by adjusting the timing of product listings in real time and combining feedback from conversion rate data, the system achieved dynamic adaptation and self-optimization of the live streaming sales scenario.

[0126] Furthermore, the e-commerce live broadcast product listing timing prediction method also includes dynamically adjusting the parameters in the weight model based on the conversion rate data and audience behavior characteristics to improve the timing prediction accuracy and conversion rate of future live broadcasts.

[0127] In the specific implementation, it is first necessary to extract core information from the conversion rate data recorded in step S105. These data include but are not limited to the actual sales volume of the product, the click rate of the audience on the product, the changes in interactive behavior (such as the changes in the number of comments and likes), and the length of time the audience stays in the live broadcast room after the product is put on the shelves. These data are matched with the comprehensive timing score of the product listing to evaluate the accuracy of the model prediction and the effectiveness of the decision. By comparing the comprehensive timing score before listing and the conversion effect after listing, the system can identify the optimization space of the current weight model on different behavioral characteristics.

[0128] When adjusting the parameters of the weight model, it is necessary to focus on analyzing the actual impact of audience behavior characteristics on conversion rate. For example, in a live broadcast, a sudden increase in the number of comments may have a significant boost to sales, indicating that the weight of interaction frequency needs to be appropriately increased. In another live broadcast, the audience's comments on a specific product are focused on a certain topic, which indicates that the characteristics of the content may be more important and their weights need to be adjusted dynamically. In this way, the weight model can be customized and optimized based on specific scenarios to ensure that the model's judgment on the importance of each type of behavior feature is more in line with actual needs.

[0129] During the dynamic adjustment process, a feedback mechanism based on historical data can be used. Specifically, the system can introduce a sliding time window to analyze the correlation between the changing trend of each audience behavior feature in recent live broadcasts and the conversion rate. For example, when a significant increase in viewing time is associated with an increase in conversion rate in past live broadcasts, the weight of participation depth can be increased; conversely, when the impact of a certain feature continues to decline, its weight can be appropriately reduced. In this way, the parameter adjustment of the weight model can smoothly respond to short-term data fluctuations while optimizing the model's predictive performance in the long term.

[0130] In order to achieve real-time and high efficiency of dynamic adjustment, the parameter adjustment process can be managed through automated algorithms. For example, based on the gradient descent method, the weight parameters are dynamically optimized to minimize the deviation between the comprehensive timing score output by the model and the actual conversion rate. This automated adjustment method can significantly improve optimization efficiency while reducing the need for manual intervention.

[0131] By dynamically adjusting the parameters of the weight model, the system is able to achieve self-learning capabilities, continuously improving its understanding and evaluation of audience behavior in each live broadcast, significantly improving the accuracy of future live broadcast timing predictions and sales conversion rates.

[0132] A second embodiment of the present application provides an electronic device, the electronic device comprising: processor; The memory is used to store a program. When the program is read and executed by the processor, it executes a method for predicting the timing of listing products in an e-commerce live broadcast provided in the first embodiment of the present application.

[0133] The third embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, a method for predicting the timing of listing products in an e-commerce live broadcast provided in the first embodiment of the present application is executed.

[0134] Although the present application is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present application. Any technical personnel in this field may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. A method for predicting the timing of listing products in e-commerce live broadcast, characterized in that: include: Acquire audience interaction data generated during the live broadcast in real time through a preset data collection device or interface, wherein the audience interaction data includes comments, likes, sharing times and viewing time sent by the audience; Preprocessing the collected audience interaction data to extract audience behavior characteristics, wherein the audience behavior characteristics include interaction frequency, participation depth, and attention content; Constructing a dynamic weight model for analyzing audience behavior characteristics, wherein the dynamic weight model is used to dynamically adjust the weight of the impact of different behavior characteristics on the decision of the time of listing; Using the dynamic weight model, combined with real-time audience behavior feature data, a comprehensive timing score is generated to evaluate the product listing potential at each time point during the live broadcast; According to the comprehensive timing score, the timing of product listing is adjusted in real time during the live broadcast, and the conversion rate data after listing is recorded to optimize the timing prediction in subsequent live broadcasts.

2. The method for predicting the timing of listing of products in live e-commerce broadcast according to claim 1, characterized in that: It also includes dynamically adjusting the parameters in the weight model based on the conversion rate data and audience behavior characteristics to improve the timing prediction accuracy and conversion rate of future live broadcasts.

3. The method for predicting the timing of listing of products in live e-commerce broadcast according to claim 1, characterized in that: The method of acquiring audience interaction data generated during the live broadcast in real time through a preset data acquisition device or interface includes: Monitor the interactive behaviors in the live broadcast room in real time, including the frequency of comments, likes, shares, and bullet comments, and use them as initial behavioral data; Through the interface connection with the live broadcast platform, the monitored interactive data is securely transmitted to the data processing unit to ensure the real-time and integrity of data collection; The interactive data transmitted to the data processing unit is pre-processed, including deduplication, timestamp addition and data classification, to generate structured data.

4. The method for predicting the timing of listing of products in live e-commerce broadcasting according to claim 1, characterized in that: The method of acquiring audience interaction data generated during the live broadcast in real time through a preset data acquisition device or interface includes: Capture audience behavior data in real time, including text interaction, emotional interaction and behavioral interaction; Label each interaction based on the type of data captured to differentiate the category and intent of the audience’s behavior; Stores tagged audience interaction data in a dynamic buffer.

5. The method for predicting the timing of listing of products in live e-commerce broadcasting according to claim 1, characterized in that: The preprocessing of the collected audience interaction data to extract audience behavior characteristics includes: De-duplicate and denoise the collected audience interaction data, filter out invalid and abnormal data, and ensure data quality; The cleaned data is classified and processed according to interaction frequency, participation depth and attention content. The interaction frequency is calculated by counting the number of interactions of each viewer in unit time; the participation depth is determined by analyzing the audience's viewing time and the diversity of participation behavior; the attention content is clustered by extracting keywords from audience comments and barrages.

6. The method for predicting the timing of listing of products in live e-commerce broadcast according to claim 1, characterized in that: The dynamic weight model for analyzing audience behavior characteristics is constructed, including: The interaction frequency, participation depth and attention content in the audience behavior characteristics are classified and processed by the feature classification module of the dynamic weight model to generate a high-frequency feature vector, a depth feature vector and a content feature vector respectively; Using the feature weight allocation module of the dynamic weight model, the feature vector generated by the feature classification unit is preliminarily weighted to generate a preliminary weighted feature value, and the weight value of each feature is dynamically adjusted according to the correlation between the preset rules and the historical audience behavior data; Through the feature association analysis module of the dynamic weight model, the preliminary weighted feature value is associated with the historical behavior feature data, the behavior feature change trend in different time periods is analyzed, and the association feature matrix is ​​generated; Using the feature fusion module of the dynamic weight model, the high-frequency features, deep features and content features in the correlation feature matrix are fused to generate a comprehensive feature value, wherein the generation of the comprehensive feature value takes into account the dynamic impact of each type of feature on the conversion rate at the current moment; The comprehensive feature value is input into the timing scoring function through the scoring generation module of the dynamic weight model, and the comprehensive timing score is calculated in combination with the real-time data during the live broadcast, wherein the comprehensive timing score is used to evaluate the product listing potential at the current time point.

7. The method for predicting the timing of listing of products in live e-commerce broadcast according to claim 6, characterized in that: The feature weight allocation module using the dynamic weight model performs preliminary weighting processing on the feature vector generated by the feature classification unit to generate preliminary weighted feature values, including: According to the following formula (1), the high-frequency feature vector, depth feature vector and content feature vector generated by the feature classification unit are preliminarily weighted by a dynamic weighting function: ； in, Features The preliminary weighted eigenvalues ​​of ; Features The original eigenvalues ​​of Features The weight value is calculated according to the following formula (2): ； in, Features Real-time trigger frequency; , , , , , , Adjust parameters for presets; Features Relevance scores in historical audience behavior data; Features rate of change in real-time audience behavior data; Relevance score Calculate according to the following formula (3): ； in, Features At the point in time The audience behavior characteristic value; For time point Behavior weight factor; is the time span of historical data; Rate of change Calculate according to the following formula (4): ； in, Features The feature value at the previous time point; To avoid division by zero errors on positive numbers; Features The feature value at the current time point; is the magnification factor; Features The historical average characteristic value of is the adjustment factor; According to the following formula (5), the standardized weighted eigenvalues ​​are generated : ； in, is the total number of features.

8. The method for predicting the timing of listing of products in live e-commerce broadcast according to claim 7, characterized in that: The feature association analysis module of the dynamic weight model performs association calculation on the preliminary weighted feature value and the historical behavior feature data, analyzes the behavior feature change trend in different time periods, and generates an association feature matrix, including: Historical behavior feature data is analyzed at preset time intervals Split into multiple time periods , in Index for time period; According to the following formula (6), the feature is calculated and In time period The strength of the association within : ； in, Features The preliminary weighted eigenvalues ​​of ; Features At the point in time The audience behavior characteristic value; is the time decay rate; For time period The central time point of Combine the correlation strength of all time periods to generate the final correlation feature matrix , whose elements Representation characteristics and Features The comprehensive correlation strength is calculated according to the following formula (7): ； in, Indicates the total number of time periods; Indicates time period The weight is calculated according to the following formula (8): ； in, Adjust parameters for weights; Indicates the index of the time period closest to the current time point of the live broadcast.

9. The method for predicting the timing of listing of products in live e-commerce broadcast according to claim 6, characterized in that: The feature fusion module using the dynamic weight model fuses the high-frequency features, deep features and content features in the associated feature matrix to generate a comprehensive feature value, including: Extract the correlation data of high-frequency features, deep features, and content features from the correlation feature matrix and use them as inputs for feature fusion, respectively, to ensure that each type of feature can independently reflect its role in audience behavior; Dynamically determine the weight of each feature type based on real-time audience behavior and historical behavior trends, ensuring that the feature categories that have a greater impact on conversion rate at the current moment have a higher weight in the fusion process; The high-frequency features, deep features and content features are weightedly fused according to the assigned weights to generate comprehensive feature values.

10. The method for predicting the timing of listing of products in live e-commerce broadcast according to claim 6, characterized in that: The scoring generation module of the dynamic weight model inputs the comprehensive feature value into the timing scoring function, and calculates the comprehensive timing score in combination with the real-time data during the live broadcast, including: The comprehensive timing score is calculated according to the following formula (9): ； in, is the number of dimensions of the comprehensive eigenvalue; The number of indicators for real-time data; Indicates Class comprehensive characteristic value; Indicates The weight of the class feature; Indicates Real-time indicator values; Indicates The weight of the real-time indicator; represents the weighted growth rate of audience behavior; and To adjust the parameters; Indicates the current live broadcast time point; The peak time point of behavior predicted by historical audience behavior data; To adjust the parameters.

Images