Live broadcast automobile marketing flow fluctuation prediction method and system

Abstract

The invention provides a live broadcast automobile marketing flow fluctuation prediction method and system, and relates to the technical field of cross-domain fusion, and the method comprises the steps: synchronously collecting data in a live broadcast platform and cross-domain extension data, and generating a platform data set and a cross-domain data set; based on the cross-domain data set, constructing an association space of the social hot events and the vehicle type attributes through a mirror image algorithm, and when the event feature intensity of the association space exceeds a preset threshold value, generating an event influence feature value; based on the event influence characteristic value and the platform data set, generating a basic flow prediction curve according to a fixed period and selecting a key point set; and processing the key point set through a convex hull algorithm to generate a convex polygon region. According to the invention, through constructing a full-link closed-loop design of data acquisition-event association-traffic prediction-parameter optimization-promotion triggering-verbal skill adjustment, accurate control of live broadcast automobile marketing traffic fluctuation and improvement of the marketing effect are realized.

Description

Technical Field

[0001] This invention relates to the field of cross-domain integration technology, and in particular to a method and system for predicting traffic fluctuations in live-streamed car marketing. Background Technology

[0002] With the deep integration of live-streaming e-commerce and the automotive industry, live-streaming has become a core channel for automotive brand promotion, model marketing, and user conversion. Industry data shows that in 2024, the transaction volume of automotive live-streaming sales increased by over 150% year-on-year. The proportion of users obtaining vehicle information, participating in interactions, and completing test drive appointments through live-streaming has significantly increased. Accurately grasping the fluctuation patterns of live-streaming traffic has become crucial for improving marketing efficiency. However, current automotive live-streaming traffic prediction and marketing decisions may face the following technical challenges:

[0003] Existing traffic prediction methods sometimes rely on data within the live streaming platform (such as the number of viewers and comments), neglecting the correlation value of cross-domain information. Cross-domain data such as public opinion dynamics in vertical automotive forums (such as user discussions on privacy configurations and price sensitivity) and the changing trends of offline 4S store inquiries are strongly correlated with live streaming traffic. However, due to heterogeneous data formats and inconsistent time granularity, they may be difficult to integrate effectively and fail to capture the linkage effect of online live streaming, cross-domain public opinion, and offline behavior. As a high-value consumer product, automobiles may be easily affected by social hot events (such as privacy and security policies and competitor promotional activities), but existing methods may lack a quantitative mechanism for the correlation between events and vehicle attributes. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method and system for predicting traffic fluctuations in live-stream car marketing. By constructing a closed-loop design that integrates data collection, event correlation, traffic prediction, parameter optimization, promotion triggering, and script adjustment, the invention achieves precise control over traffic fluctuations in live-stream car marketing and improves marketing effectiveness.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, a method for predicting traffic fluctuations in live-stream car marketing, the method comprising:

[0007] Step 1: Synchronously collect data within the live streaming platform and cross-domain extended data to generate platform datasets and cross-domain datasets;

[0008] Step 2: Based on the cross-domain dataset, construct the association space between social hot topics and vehicle attributes using the mirroring algorithm. When the event feature strength in the association space exceeds a preset threshold, generate event impact feature values.

[0009] Step 3: Based on the event impact feature values ​​and platform dataset, generate basic traffic prediction curves at fixed intervals and select key point sets; process the key point sets using the convex hull algorithm to generate convex polygon regions;

[0010] Step 4: When segmenting the convex polygon region into grid cells, simultaneously calculate the traffic change gradient of each grid cell to generate a gradient distribution matrix and identify the vehicle type; if it is a high-end vehicle, call the mirror algorithm based on the historical feature library, establish a privacy label weight mapping relationship based on the traffic change gradient, and generate optimized weight parameters.

[0011] Step 5: Generate grid adjustment feature quantity based on optimized weight parameters and gradient distribution matrix; merge event impact feature value and grid adjustment feature quantity to generate comprehensive correction parameter, and correct the basic traffic prediction curve to generate corrected traffic curve; when the corrected traffic curve shows a downward trend in traffic and the optimized weight parameter meets the preset threshold, pre-generate promotion trigger instruction;

[0012] Step 6: Execute the promotion trigger instruction based on the pre-generated promotion trigger instruction and cross-domain dataset; at the same time, process the public opinion correlation between the cross-domain dataset and the live broadcast room comments through the mirroring algorithm, and execute the wording optimization instruction when the correlation exceeds the preset threshold.

[0013] Secondly, a live-streaming car marketing traffic fluctuation prediction system includes:

[0014] The acquisition module is used to simultaneously acquire data within the live streaming platform and cross-domain extended data, generating platform datasets and cross-domain datasets;

[0015] The association module is used to construct an association space between social hot topics and vehicle attributes based on cross-domain datasets and through mirroring algorithms. When the intensity of event features in the association space exceeds a preset threshold, event impact feature values ​​are generated.

[0016] The convex hull algorithm module is used to generate basic traffic prediction curves and select key point sets at fixed intervals based on event impact feature values ​​and platform datasets; and to generate convex polygon regions by processing the key point sets through the convex hull algorithm.

[0017] The optimized weight module is used to simultaneously calculate the gradient distribution matrix of the traffic change gradient of each grid cell when segmenting the convex polygon region into grid cells, and to identify the vehicle type; if it is a high-end vehicle, it calls the mirror algorithm based on the historical feature library, establishes a privacy label weight mapping relationship based on the traffic change gradient, and generates optimized weight parameters.

[0018] The correction module is used to generate grid adjustment feature quantities based on optimized weight parameters and gradient distribution matrix; it integrates event impact feature values ​​and grid adjustment feature quantities to generate comprehensive correction parameters, and corrects the basic traffic prediction curve to generate a corrected traffic curve; when the corrected traffic curve shows a downward trend in traffic and the optimized weight parameters meet the preset threshold, a promotion trigger instruction is pre-generated.

[0019] The optimization instruction module is used to execute promotion trigger instructions based on pre-generated promotion trigger instructions and cross-domain datasets; at the same time, it uses a mirroring algorithm to process the correlation between cross-domain datasets and live stream comments, and executes the wording optimization instruction when the correlation exceeds a preset threshold.

[0020] Thirdly, a computing device includes:

[0021] One or more processors;

[0022] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0023] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0024] The above-described solution of the present invention has at least the following beneficial effects:

[0025] By simultaneously collecting data from the live streaming platform and expanding data across domains, the limitations of single-source data within the platform are broken. Simultaneously, timeline alignment and feature fusion ensure consistency across multiple data sources in the spatiotemporal dimension. A mirroring algorithm is used to construct a correlation space between social hot topics and vehicle attributes, transforming abstract social hot topics into quantifiable event impact features, intuitively reflecting the dynamic impact of events on traffic for different vehicle models. For high-end vehicle privacy-sensitive periods, the convex polygon region is refined by adding auxiliary connection points, improving the accuracy of depicting traffic fluctuations during privacy-sensitive periods. For price-sensitive periods for economy vehicles, connection paths are simplified to reduce computation and optimize efficiency. This combination of precise regions and efficient paths ensures both the prediction depth of core scenarios for high-end vehicles and avoids resource waste. Based on traffic change gradients and a historical high-end vehicle feature database, the system dynamically adjusts privacy tag weight parameters using a mirroring algorithm: increasing the weight of high-attention tags when traffic rises and decreasing the weight when traffic falls, responding in real-time to fluctuations in user interest. A corrected traffic curve is generated by comprehensively adjusting parameters. When a continuous decline in traffic is detected and users show high attention to core tags, promotional trigger instructions are automatically pre-generated, achieving a closed loop of traffic alerts and precise promotions. While executing promotional instructions, the system correlates cross-domain public opinion with livestream comments using a mirroring algorithm. When the correlation exceeds a threshold, optimized script instructions are automatically generated and pushed in real-time via the host's teleprompter, reducing information asymmetry for users and improving interaction satisfaction and conversion efficiency. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a method for predicting traffic fluctuations in live-stream car marketing, provided by an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of a live-streaming car marketing traffic fluctuation prediction system provided by an embodiment of the present invention. Detailed Implementation

[0028] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0029] like Figure 1 As shown in the figure, an embodiment of the present invention proposes a method for predicting traffic fluctuations in live-stream car marketing, the method comprising the following steps:

[0030] Step 1: Synchronously collect data within the live streaming platform and cross-domain extended data to generate platform datasets and cross-domain datasets;

[0031] Step 2: Based on the cross-domain dataset, construct the association space between social hot topics and vehicle attributes using the mirroring algorithm. When the event feature strength in the association space exceeds a preset threshold, generate event impact feature values.

[0032] Step 3: Based on the event impact feature values ​​and platform dataset, generate basic traffic prediction curves at fixed intervals and select key point sets; process the key point sets using the convex hull algorithm to generate convex polygon regions;

[0033] Step 4: When segmenting the convex polygon region into grid cells, simultaneously calculate the traffic change gradient of each grid cell to generate a gradient distribution matrix and identify the vehicle type; if it is a high-end vehicle, call the mirror algorithm based on the historical feature library, establish a privacy label weight mapping relationship based on the traffic change gradient, and generate optimized weight parameters.

[0034] Step 5: Generate grid adjustment feature quantity based on optimized weight parameters and gradient distribution matrix; merge event impact feature value and grid adjustment feature quantity to generate comprehensive correction parameter, and correct the basic traffic prediction curve to generate corrected traffic curve; when the corrected traffic curve shows a downward trend in traffic and the optimized weight parameter meets the preset threshold, pre-generate promotion trigger instruction;

[0035] Step 6: Execute the promotion trigger instruction based on the pre-generated promotion trigger instruction and cross-domain dataset; at the same time, process the public opinion correlation between the cross-domain dataset and the live broadcast room comments through the mirroring algorithm, and execute the wording optimization instruction when the correlation exceeds the preset threshold.

[0036] In this embodiment of the invention, data within the live streaming platform and cross-domain extended data are collected simultaneously to ensure the integrity of the platform dataset and the cross-domain dataset. A correlation space between social hot topics and vehicle attributes is constructed using a mirroring algorithm, accurately capturing the impact of events on vehicle models and making the generation of event impact feature values ​​more targeted. A basic traffic prediction curve is generated based on the event impact feature values ​​and the platform dataset. Combined with the processing of key point sets and convex polygon regions, a refined characterization of traffic fluctuation trends is achieved. Segmenting the convex polygon region into grid cells and calculating the traffic change gradient accurately reflects the traffic dynamics in different time periods and regions. Further improvement in traffic prediction accuracy is achieved through vehicle type identification and weight parameter optimization. A comprehensive correction parameter is generated by fusing event impact feature values ​​and grid adjustment feature quantities to correct the basic traffic prediction curve. When the correction curve shows a downward traffic trend and meets the conditions, a promotional trigger instruction is pre-generated, enabling timely responses to traffic loss risks in promotional decisions. While executing the promotional trigger instruction, the correlation between the cross-domain dataset and live stream comments is processed using a mirroring algorithm. When the correlation exceeds a threshold, a script optimization instruction is executed, improving the quality of interaction with users and enhancing the attractiveness and persuasiveness of promotional activities.

[0037] In a preferred embodiment of the present invention, step 1 above, which involves simultaneously collecting data within the live streaming platform and cross-domain extended data to generate a platform dataset and a cross-domain dataset, may include:

[0038] Step 11: Collect platform dynamic data in real time from the pre-live broadcast period to the end of the live broadcast to obtain raw platform data;

[0039] Step 12: Based on the original platform data, simultaneously trigger cross-domain data collection of the target model's public opinion keyword set in the vertical automotive forum and the consultation volume trend data of offline sales channels to obtain the original cross-domain data.

[0040] Step 13: Based on the original cross-domain data and the original platform data, align the original platform data according to the live broadcast timeline to generate time-aligned platform data; based on the time-aligned platform data, extract the interaction rate-conversion rate correlation matrix; synchronize the original cross-domain data with the aligned platform timeline to generate time-aligned cross-domain data.

[0041] Step 14: Perform feature encoding on time-aligned platform data to generate a platform dataset containing a peak viewing distribution matrix, an interaction conversion correlation tensor, and a competitor traffic time-series feature vector; perform semantic aggregation on time-aligned cross-domain data to generate a cross-domain dataset containing public opinion keyword weight distribution data and an offline consultation volume gradient sequence.

[0042] In this embodiment of the invention, real-time data collection throughout the entire period ensures the integrity and timeliness of the original platform data, providing accurate basic data support for subsequent analysis; dynamic cross-domain data collection based on the target vehicle model enables the association between online platforms and offline channels and public opinion data, avoiding invalid data and improving data targeting; timeline alignment eliminates the time granularity differences of multi-source data, laying the foundation for cross-domain data fusion and correlation analysis, and ensuring the effectiveness of data comparison; feature encoding and semantic aggregation transform the original data into a structured dataset, improving data processing efficiency.

[0043] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0044] Step 11 above involves setting up a real-time data collection trigger mechanism from the pre-live broadcast warm-up period (e.g., 72 hours before the live broadcast) to 24 hours after the live broadcast ends; and collecting raw data such as the number of viewers, cumulative viewing time, number of comments, number of likes, number of shares, and number of clicks on product links at a second-level frequency through the live broadcast platform's backend interface, and storing it as a raw platform data sequence with timestamps.

[0045] In step 12 above, when the target vehicle model name appears in the original platform data (e.g., identified through keyword matching), the cross-domain data acquisition unit is automatically activated.

[0046] Collect data from vertical automotive forums: crawl posts and comments containing the target car model, extract public opinion keywords (such as "range", "price", "configuration", etc.), count the frequency of each keyword, and form a public opinion keyword set.

[0047] Collect offline channel data: Connect to the offline 4S store management platform to count the number of in-store inquiries and telephone inquiries for the target model on an hourly basis, generate raw trend data of inquiries changing over time, and establish a correlation mapping with the platform data.

[0048] Step 13 above uses the live broadcast timeline (e.g., the broadcast start time is 0 o'clock, and the time period is divided into minutes) as the basis to aggregate the original platform data according to the same time granularity (e.g., per minute) to generate time-aligned platform data (e.g., "the number of viewers in the 10th minute = the average value of the real-time data in that minute").

[0049] Extracting the interaction rate-conversion rate correlation matrix involves using time-aligned platform data to calculate the ratio of "number of interaction behaviors (comments + likes)" to "number of conversion behaviors (product clicks)" within each time period, and constructing a matrix by time period × behavior type (rows represent time periods, and columns represent interaction-conversion correlation values).

[0050] The original cross-domain data (forum keyword frequency, offline consultation volume) is interpolated and aligned according to the same granularity (minute level) of the platform's timeline. For example, the hourly offline consultation volume is broken down into minutes to generate time-aligned cross-domain data.

[0051] Step 14 above, platform dataset feature encoding:

[0052] The peak viewing distribution matrix is ​​platform data aligned to time, which counts the peak number of viewers for each time period and encodes it in a two-dimensional table format of "time period × peak value", marking the specific time point when the peak occurred.

[0053] The interaction-conversion correlation tensor combines the interaction rate-conversion rate correlation matrix with the time dimension, integrating it into a three-dimensional structure of "time period × interaction type × conversion efficiency" to form tensor data.

[0054] The competitor traffic time series feature vector is generated by collecting the number of viewers in the competitor's live broadcast room during the same period, aligning it according to the time axis, and then generating a vector to compare with the target model's traffic (such as the "target model traffic - competitor traffic" difference sequence).

[0055] Semantic aggregation across domain datasets:

[0056] The public opinion keyword weight distribution data is a weight distribution chart that calculates the weight percentage of each keyword in cross-domain data that is aligned with time. The data is based on the frequency of occurrence of public opinion keywords (the higher the frequency, the greater the weight).

[0057] The offline consultation volume gradient sequence is generated by calculating the difference in offline consultation volume between adjacent time periods (current time period consultation volume - previous time period consultation volume), and generating a gradient sequence that reflects the trend of increase or decrease in consultation volume. Positive values ​​indicate increase, and negative values ​​indicate decrease.

[0058] In a preferred embodiment of the present invention, step 2 above, based on a cross-domain dataset, constructs an association space between social hot topics and vehicle attributes using a mirroring algorithm. When the event feature strength in the association space exceeds a preset threshold, event impact feature values ​​are generated, which may include:

[0059] Step 21: Based on the cross-domain dataset, analyze the public opinion keyword weight distribution data and extract the high-frequency keyword set; standardize the offline consultation volume gradient sequence and generate a consultation volume change vector to obtain the preprocessed cross-domain dataset.

[0060] Step 22: Based on the preprocessed cross-domain dataset, the high-frequency keyword set is mapped to the vehicle attribute space using the mirroring algorithm. The consultation volume change vector is then subjected to tensor product operation with the vehicle attribute space to obtain the social hotspot-vehicle attribute association space.

[0061] Step 23: Calculate the feature vector magnitude of each social hotspot event in the social hotspot-vehicle attribute association space, normalize the magnitude, and generate a sequence of event feature intensity values.

[0062] Step 24: When any intensity value in the event feature intensity value sequence exceeds a preset dynamic threshold, feature value generation is triggered and the intensity value exceeding the threshold is converted into event impact feature value.

[0063] In this embodiment of the invention, by filtering high-frequency keywords and standardizing consultation volume gradients, low-relevance data is eliminated and data scales are unified, making the preprocessed cross-domain data more focused on core public opinion and trend characteristics. A mirroring algorithm is used to achieve accurate mapping between hot keywords and vehicle attributes, and a correlation space is constructed by combining tensor product operations to intuitively quantify the dynamic relationship between the two. Through feature vector magnitude calculation and normalization, the abstract impact of social hotspots is transformed into comparable quantitative intensity values, making the impact of different events horizontally comparable. Based on dynamic threshold triggering feature value generation, it ensures that impactful events are accurately captured, and dynamically adjusts the threshold to adapt to different live streaming scenarios, improving the targeting and effectiveness of event impact feature values.

[0064] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0065] In step 21 above, the weight distribution data of public opinion keywords in the cross-domain dataset is extracted, and the weight ratio of each keyword is calculated according to the time period (e.g., the weight ratio of "price" in time period 1 is 30%). Keywords with a weight ratio exceeding the preset threshold (e.g., 20%) are selected, and these frequently occurring keywords (e.g., "range", "intelligent driving", "preferential policies") are summarized to form a high-frequency keyword set.

[0066] Take the gradient sequence of offline consultation volume from the cross-domain dataset (containing positive and negative values, reflecting the increasing and decreasing trend), and use the extreme value scaling method to uniformly map the gradient values ​​to the 0-1 interval (e.g., the maximum gradient value corresponds to 1, and the minimum gradient value corresponds to 0), eliminate the difference in the scale of data in different time periods, and generate a standardized consultation volume change vector (each time period corresponds to a standardized value).

[0067] By associating the set of high-frequency keywords with the vector of changes in consultation volume, a preprocessed cross-domain dataset containing a three-dimensional relationship of "keyword-time period-standardized consultation volume" is formed.

[0068] Step 22 above maps high-frequency keywords to vehicle attribute space: The mirroring algorithm is called to match the set of high-frequency keywords generated in step 21 with the preset vehicle attribute space (including the core attribute dimensions of the vehicle, such as price range, power type, configuration level, target user group, etc.) - for example, "preferential policy" corresponds to "price attribute", "intelligent driving" corresponds to "configuration level attribute", so that each keyword corresponds to a specific vehicle attribute dimension, and the mapping from keyword to attribute space is completed.

[0069] The mapped keyword attribute vector and the consultation volume change vector are subjected to tensor product operation, that is, the data is integrated according to the three-dimensional structure of "keyword corresponding attribute × consultation volume change trend × time period" (such as the value corresponding to the consultation volume growth trend of "price attribute" in time period 1), forming a social hot spot - vehicle model attribute association space, which intuitively presents the dynamic relationship between social hot spot keywords and vehicle model attributes.

[0070] In step 23 above, in the social hotspot-vehicle model attribute association space, the associated data corresponding to each social hotspot event (such as "a celebrity recommends the same car model") is transformed into a feature vector (containing the keywords involved in the event, the associated vehicle model attributes, the trend of consultation volume, etc.), and the magnitude of the vector is calculated (i.e. the square root of the sum of the squares of the values ​​of each dimension of the vector, reflecting the overall strength of the event's impact).

[0071] Normalize the feature vector magnitudes of all social hot topics by the maximum value (i.e., divide each magnitude by the maximum magnitude) to obtain the event feature intensity values ​​in the 0-1 interval. Arrange the events in chronological order to generate a sequence of event feature intensity values.

[0072] In step 24 above, based on the intensity range of "events significantly altering traffic trends" in historical data, a dynamic threshold is set (e.g., a threshold of 0.6 for non-promotional periods and 0.4 for promotional periods, adaptively adjusted according to the live streaming stage).

[0073] The intensity value sequence is monitored in real time. When any intensity value exceeds the preset dynamic threshold for the corresponding time period, the feature value generation mechanism is automatically triggered to directly convert the intensity value into an event impact feature value (the normalized value is retained for subsequent traffic prediction correction).

[0074] In a preferred embodiment of the present invention, step 3 above, which involves generating a basic traffic prediction curve and selecting a key point set based on event impact feature values ​​and platform datasets at fixed intervals, and generating a convex polygon region by processing the key point set using a convex hull algorithm, may include:

[0075] Step 31: Inject the event impact feature values ​​into the interaction transformation association tensor of the platform dataset, and generate a spatiotemporal augmented dataset based on the fused tensor;

[0076] Step 32: Based on the spatiotemporal augmentation dataset, perform time series prediction at fixed intervals to generate a basic traffic prediction curve; correct the shape characteristics of the basic traffic prediction curve based on the historical high-end vehicle feature library, and automatically select a set of key points, where: peak points correspond to promotional sensitive periods, trough points correspond to periods of traffic loss risk, and inflection points correspond to periods of public opinion change response.

[0077] Step 33: Select boundary extreme points in the key point set using the convex hull algorithm, and connect them in chronological order to generate an initial convex polygon region; during the connection process, add auxiliary connection points to the key point set corresponding to the privacy label period of high-end models to improve the region accuracy, and simplify the connection path to optimize the calculation efficiency for the key point set corresponding to the price sensitive period of economy models, thereby generating a convex polygon region covering the core traffic fluctuation area.

[0078] In this embodiment of the invention, the spatiotemporal correlation of the dataset is enhanced by fusing event features with interactive conversion data, solving the problems of lack of correlation and incomplete information in single-platform data, and providing more comprehensive feature support for accurate prediction; fixed-period prediction combined with historical feature correction improves the accuracy of the basic traffic curve; classification and identification of key point sets make the key periods of traffic fluctuation clearer; and differentiated region optimization based on vehicle type ensures the accuracy of characterizing the privacy periods of high-end vehicles while improving the computational efficiency of economy vehicles. The resulting convex polygon region can accurately cover the core traffic fluctuation area, laying a reliable foundation for grid segmentation and traffic analysis.

[0079] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0080] In step 31 above, the interaction conversion association tensor of the platform dataset is extracted. This tensor contains feature data in three dimensions: time period, interaction type, and conversion type. The event impact feature values ​​are aligned with the interaction conversion association tensor along the time dimension. The tensor is then injected through dimension expansion (such as adding a "event impact intensity" dimension) so that the tensor contains both the original interaction conversion features and the event impact features.

[0081] The injected tensors are integrated in a spatiotemporal dimension, and the data blocks are divided according to the live broadcast timeline and user interaction scenarios (such as consultation scenarios and order placement scenarios) to generate a spatiotemporally enhanced dataset.

[0082] Step 32 above divides the entire live broadcast into equal-length time windows (e.g., each 30-minute period), with each period serving as a prediction unit, forming a continuous time series.

[0083] For each prediction period, using data from the previous N periods (e.g., the previous 3 periods) as a historical window, the following features are extracted:

[0084] Traffic statistics characteristics: average number of viewers within the window, maximum / minimum number of viewers, and standard deviation of viewers (reflecting the fluctuation range).

[0085] Interactive conversion characteristics: Trends in in-window interaction rate (number of comments / number of viewers) and conversion rate (number of clicks / number of viewers).

[0086] Event impact characteristics: The cumulative intensity and slope of change (upward / downward trend) of the event impact characteristic values ​​within the window.

[0087] Based on the extracted historical features and the event impact features of the current period (such as the weight of public opinion keywords and the gradient of offline consultation volume), a basic traffic prediction curve for the future period is generated.

[0088] The historical high-end vehicle feature database is a structured feature set formed by mining, extracting, and verifying a large amount of historical high-end vehicle live streaming data. It is used to support the prediction and correction of current live streaming traffic. Its establishment process is as follows:

[0089] We collected live stream records for all high-end car models (such as sedans / SUVs priced at ≥ 500,000 RMB) from the past 1-2 years, covering different brands (such as Mercedes-Benz, BMW, and Tesla's high-end series) and different live stream scenarios (such as new car launches and promotional events); real-time viewership, interaction data (comments / likes / shares), conversion data (number of clicks for test drives / inquiries), user dwell time, and bullet screen keywords (such as "privacy glass" and "rear seat sound insulation") for each live stream; and public opinion on vertical automotive forums corresponding to the live stream time (such as post keywords in the high-end car section of "Autohome"), offline 4S store inquiries (such as "percentage of customers inquiring about privacy configurations at stores"), and social hot topics (such as "the introduction of privacy and security policies for high-end car models").

[0090] The following core features were extracted from the collected historical data (categorized by business scenario): Traffic fluctuation characteristics: the ratio of peak traffic during privacy-tagged periods (such as discussions on "privacy configuration") to that during normal periods (historical average is 1.3:1); the duration of peak traffic (average 28 minutes for high-end models, 10 minutes longer than for economy models); and the rate of traffic decline when negative public opinion (such as "privacy leakage risk") occurs (historical average is -30 people / minute).

[0091] Time-sensitive characteristics: Historical data shows that 42% of high-end users watch the content between 20:00 and 21:30 on weekdays (after get off work); users stay for an average of 19 minutes, with 70% of that time concentrated in the "configuration explanation" section.

[0092] Public opinion correlation characteristics: After the keyword "privacy configuration" appears, the traffic peak appears with an average delay of 8 minutes (users need time to digest the information); for every 10% increase in the weight of the keyword "smart privacy", the traffic peak increases by 12% (historical statistics).

[0093] Conversion characteristics: The conversion rate (number of clicks for inquiry / number of viewers) during the privacy label period was 17% higher than that during normal periods; the linkage between offline inquiries and online traffic: for every 5 additional offline inquiries, the peak online traffic increased by an average of 8% (users tend to "learn online + verify offline").

[0094] Extreme data (such as a sudden drop in traffic due to live streaming failure) is removed. Abnormal feature values ​​exceeding "mean ± 3 standard deviations" are identified and deleted using the IQR (interquartile range) method. Features are standardized to the format of "feature name-time granularity-statistic value", such as "peak traffic ratio during privacy period - 30-minute cycle - 1.3" and "keyword response delay - minutes - 8", to ensure that data from different historical live streams can be directly compared.

[0095] Select 3-5 typical historical live streams (such as a live stream for the launch of a high-end car model), and use the extracted features to back-engineer the traffic curve to verify the accuracy of the features; for example, if the feature library shows "privacy keyword weight increased by 10% → traffic peak +12%", then backtrack the actual data of that live stream. If the deviation is ≤5%, the feature is effective; test the universality of features in different brands / scenarios (such as whether "high attention to privacy configuration" applies to all high-end car models), and separately label brand-specific features (such as the traffic association of a brand's "starry sky canopy" configuration).

[0096] Add nearly three months of live streaming data for high-end models each quarter, repeat steps 2-4, and update the feature library (such as adding privacy features for "new energy high-end models"); when new user behavior trends emerge (such as the impact of "metaverse test drive" on traffic), supplement new feature dimensions to ensure the timeliness of the library.

[0097] Recording the traffic curve patterns of live streams featuring high-end car models in history, it can be found that the peak traffic during privacy-themed periods (such as when discussing "privacy glass" or "sound insulation features") is typically 20%-30% higher than during normal periods, and the fluctuation frequency is lower (lasts longer).

[0098] Traffic patterns during specific time periods, such as high-end models experiencing a 50% higher traffic peak between 8-10 PM on weekdays compared to midday, while economy models show no significant difference; recording the response pattern of the traffic curve when specific public opinion keywords related to high-end models (such as "luxury" and "technological features") appear (e.g., for every 10% increase in keyword weight, the traffic peak appears 15 minutes later); and the resilience of the high-end model traffic curve when competitors launch similar models (e.g., the decline in high-end model traffic is 15% lower than that of economy model traffic).

[0099] The generated basic traffic prediction curve is compared with the typical curve in the historical high-end vehicle feature library, and the similarity is calculated (e.g., by using the Dynamic Time Warping (DTW) algorithm). If it is found that the peak value of the current curve in the privacy label period is lower than the historical average, it is corrected in the following ways: the predicted traffic value corresponding to the privacy label period is increased by 15%-20%; the rate of change of the derivative of the curve in this period is reduced to make the peak value closer to the historical typical shape.

[0100] Calculate the first derivative of the curve. Points where the derivative is 0 are marked as peaks (highest traffic points) or troughs (lowest traffic points). Points where the derivative changes abruptly are marked as inflection points (points where traffic trends change). These are then categorized into "promotion-sensitive periods (peaks), traffic loss-risk periods (troughs), and public opinion change response periods (inflection points)" to form a set of key points. Promotion-sensitive periods (peaks) are the periods with the highest traffic, typically corresponding to the time when audience activity and purchasing intention are strongest, making them suitable for promotional activities (such as limited-time discounts). Traffic loss-risk periods (troughs) are the periods with the lowest traffic, potentially indicating declining audience interest or competitor diversion, requiring early warning and adjustments to the live stream content. Public opinion change response periods (inflection points) are points where traffic trends suddenly change, usually related to hot events or negative public opinion, requiring timely adjustments to the wording or addition of interactive elements.

[0101] In step 33 above, the key point set is sorted according to the amplitude of traffic fluctuation, and the boundary feature points of the fluctuation (such as the first 20% of peaks / troughs / inflection points) are selected. These points are then connected by the convex hull algorithm to form an initial convex polygon region covering the main range of traffic fluctuation.

[0102] For high-end models: Identify the set of key points corresponding to the privacy label period (such as peaks / inflection points associated with public opinion on "privacy configuration"), insert interpolation auxiliary points within the time interval of adjacent key points, and increase the density of regional vertices to refine the characterization of traffic fluctuations during privacy periods.

[0103] For economy car models: Identify key point sets corresponding to price-sensitive periods (such as troughs / inflection points associated with "promotional activities"), merge adjacent non-extreme original sampling points, and simplify connection paths (such as replacing broken lines with straight lines) to reduce computational load.

[0104] The adjusted key point set is integrated and reconnected to form the final convex polygon region, ensuring coverage of the core traffic fluctuation area (e.g., containing more than 80% of the key point set).

[0105] In a preferred embodiment of the present invention, in step 4 above, when segmenting the convex polygonal region into grid cells, the gradient distribution matrix is ​​generated by simultaneously calculating the traffic flow change gradient of each grid cell and identifying the vehicle type; if it is a high-end vehicle, the mirror algorithm is called based on the historical feature library to establish a privacy label weight mapping relationship according to the traffic flow change gradient and generate optimized weight parameters, which may include:

[0106] Step 41: Based on the convex polygon region, divide it into grid cells along the time axis, and extract the flow data and time period information contained in each grid cell to generate grid cell data; simultaneously calculate the flow change rate of each grid cell to generate a gradient distribution matrix;

[0107] Step 42: Based on the grid cell data and gradient distribution matrix, extract user profile features within the grid cells, including user query behavior and interaction frequency regarding vehicle configuration; call the historical high-end vehicle feature library for feature comparison, count the frequency of privacy configuration queries, and when the frequency is greater than the benchmark value in the historical high-end vehicle feature library, mark it as a high-end vehicle; otherwise, mark it as a Volkswagen vehicle to obtain the vehicle identification.

[0108] Step 43: When the vehicle model is identified as a high-end vehicle, the gradient distribution matrix is ​​invoked and a privacy label weight mapping relationship is established based on the historical high-end vehicle feature library through the mirroring algorithm. That is, the gradient change rate is used as the independent variable, the historical weight of the privacy label in the feature library is used as the base value, and the weight value is dynamically adjusted according to the gradient change direction. When the gradient increases, the weight increase is increased, and when the gradient decreases, the weight decrease is applied to generate optimized weight parameters.

[0109] In this embodiment of the invention, continuous traffic fluctuation regions are discretized through grid cell segmentation, making traffic change analysis more refined; the gradient distribution matrix is ​​calculated simultaneously to intuitively quantify the traffic change rate at each time period, providing a data foundation for subsequent weight adjustment; vehicle type is distinguished based on user query behavior for privacy configurations, aligning with the high level of attention paid to privacy attributes by users of high-end vehicles, making vehicle identification more accurate; the objectivity of classification is improved by comparing with the benchmark values ​​of the historical feature database; and the privacy label weights are dynamically adjusted in conjunction with the traffic gradient, enabling the weight parameters to adapt to traffic change trends in real time.

[0110] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0111] In step 41 above, the time axis (horizontal axis) and flow rate axis (vertical axis) of the convex polygon region are divided at equal intervals to form multiple rectangular grid units (e.g., each grid unit corresponds to a 5-minute time interval + 50-person flow rate interval), and each grid unit corresponds to a specific spatiotemporal range.

[0112] The data includes traffic data (such as the average number of viewers and the maximum traffic value during the time period) and time period information (such as start time and duration) within each grid cell. This data is stored in the format of "grid number-traffic data-time period" to generate grid cell data.

[0113] For adjacent grid cells, calculate the difference in flow rate (current grid flow rate - previous grid flow rate), and then divide by the time interval (e.g., 5 minutes) to obtain the flow rate change rate of each grid cell (positive value indicates an increase, negative value indicates a decrease); arrange the rates of all grid cells according to their positions to form a gradient distribution matrix (rows represent time order, columns represent flow ranges, and matrix elements are the change rates of the corresponding grid cells).

[0114] Step 42 above extracts user behavior features from the grid cell data:

[0115] Configuration query behavior: Statistical analysis of the frequency of vehicle configuration keywords (such as privacy features like "privacy glass", "encrypted storage", and "sound insulation cotton", and common features like "price" and "fuel consumption") mentioned by users in comments and bullet screens within each grid cell.

[0116] Interaction frequency: Count the number of comments, likes and shares of users in each grid cell, and calculate the interaction density (total number of interactions / number of viewers in the grid).

[0117] Access the historical high-end vehicle feature database and extract the "privacy configuration query baseline value" (e.g., in the historical high-end vehicle live broadcast, the average query frequency of privacy configuration is 15 times per grid).

[0118] The total query frequency of privacy configuration in the current grid cell data is counted. If the frequency is greater than the baseline value (e.g., 20 times > 15 times), the vehicle model is marked as "high-end vehicle"; otherwise, it is marked as "mass vehicle" and a vehicle model identifier is generated.

[0119] In step 43 above, two types of key grid units are selected from the gradient distribution matrix: one is the grid marked as "high-end models", and the other is the time period when users frequently query privacy configurations (such as "privacy glass", "encrypted storage", "rear sound insulation" etc.) within the grid. For these grid units, the specific characteristics of their traffic change gradient are extracted, that is, the rate of traffic change per unit time (such as +8 people / minute indicates an increase in traffic, and -5 people / minute indicates a decrease in traffic).

[0120] From the historical high-end vehicle feature database, the privacy tags (such as "privacy glass" and "encrypted storage") involved in the current live stream are matched, and the corresponding "historical base weights" are extracted. These weights are benchmark values ​​calculated based on a large amount of historical high-end vehicle live stream data, representing the long-term attention paid by users to different privacy tags in the past. For example, "privacy glass" is frequently mentioned by high-end users in historical live streams, so its base weight is set to 0.6 (that is, the historical average impact of this tag on traffic fluctuations is 60%); "encrypted storage" is a characteristic focus of new energy high-end vehicles, so its base weight is set to 0.5 (historical average impact is 50%).

[0121] Based on the gradient of flow change in grid cells, historical baseline weights are adjusted in real time, in two scenarios: When the gradient increases (rate is positive): If the rate of change in traffic is positive (e.g., +8 people / minute), it indicates that current users' attention to the privacy label is increasing. At this time, the weight is increased according to the "gradient increase magnitude": for every 2 people / minute increase in rate, the weight increases by 5% from its historical base. For example, the historical base weight for "Privacy Glass" is 0.6, and the current gradient rate is +8 people / minute (i.e., 4 more "2 people / minute" units than the base rate). The weight adjustment is: base weight 0.6 + (4 × 5% × 0.6) = 0.6 + 0.12 = 0.72. This is actually a proportional cumulative increase, meaning every 2 people / minute corresponds to a 5% increase, and 8 people / minute corresponds to 4 increase units, for a total increase of 4 × 5% = 20%. Therefore, the adjusted weight is 0.6 × (1 + 20%) = 0.72.

[0123] When the gradient decreases (the rate is negative): If the rate of change in traffic is negative (e.g., -6 people / minute), it indicates that users' attention to the privacy label is decreasing. In this case, the weight is reduced according to the "gradient decrease magnitude": for every 2 people / minute decrease in the rate, the weight is reduced by 3% from its historical value. For example, the historical base weight for "encrypted storage" is 0.5, and the current gradient rate is -6 people / minute (i.e., a decrease of 3 units more than the base rate of 2 people / minute), resulting in a total decrease of 3 × 3% = 9%. Therefore, the adjusted weight is 0.5 × (1 - 9%) = 0.455 (approximately 0.46).

[0124] The weights of all privacy tags after gradient adjustment are integrated to form a set of optimized weight parameters that adapt to the current traffic trend. For example, the weight of "privacy glass" is adjusted to 0.72, "encrypted storage" to 0.46, and other privacy tags (such as "back row soundproofing") are adjusted according to the same logic to form the final optimized weight parameters.

[0125] In a preferred embodiment of the present invention, step 5 above, which generates grid adjustment feature quantities based on optimized weight parameters and gradient distribution matrix; integrates event impact feature values ​​and grid adjustment feature quantities to generate comprehensive correction parameters, and corrects the basic traffic prediction curve to generate a corrected traffic curve; when the corrected traffic curve shows a downward trend in traffic and the optimized weight parameters meet a preset threshold, a promotion trigger instruction is pre-generated, which may include:

[0126] Step 51: Based on the optimized weight parameters and gradient distribution matrix, extract the rate of change feature through the spatiotemporal characteristics of the gradient distribution matrix, and perform weighted aggregation on the rate of change feature based on the optimized weight parameters to obtain the grid adjustment feature quantity.

[0127] Step 52: Based on the mesh adjustment feature quantity and the event impact feature value, map the event impact feature value to the preset feature space, and perform tensor fusion with the mesh adjustment feature quantity to obtain the comprehensive correction parameter;

[0128] Step 53: Inject the comprehensive correction parameters into the fluctuation range of the basic flow prediction curve to generate the corrected flow curve;

[0129] Step 54: When it is detected that the flow rate decreases for two consecutive cycles in the corrected flow rate curve is greater than the preset range, and the verification optimization weight parameter is greater than the vehicle model sensitivity threshold, a promotion trigger instruction is pre-generated.

[0130] In this embodiment of the invention, weighted aggregation combines optimized weight parameters with gradient features, making the grid adjustment feature quantity more focused on high-attention privacy tag-related grids, thus improving the feature's targeting of high-end vehicle user behavior; merging event impacts and grid features in a unified feature space avoids the limitations of single features, allowing the comprehensive correction parameters to simultaneously cover the impact of external events and internal traffic changes, enhancing the comprehensiveness of the correction; targeted injection of correction parameters adjusts the curve fluctuation range, solving the problem of the basic prediction curve ignoring the comprehensive impact of multiple factors, making the corrected traffic curve more closely match the actual traffic change pattern, thus improving prediction accuracy; triggering promotional instructions based on dual conditions of traffic trend and weight threshold ensures timely intervention when traffic declines, and also filters out high-end user groups sensitive to privacy tags through weight thresholds, avoiding ineffective promotions and improving the utilization rate of marketing resources.

[0131] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0132] Step 51 above extracts two types of rate of change features from the gradient distribution matrix:

[0133] Time dimension characteristics: Statistical analysis of the average traffic change rate of grid units in each time period according to the live broadcast timeline (e.g., every 30 minutes is a segment) and the consistency of the change trend (e.g., the proportion of three consecutive grids showing an upward trend).

[0134] Spatial dimension characteristics: Statistical analysis of the gradient standard deviation within the same interval (reflecting the stability of flow fluctuations in that interval) and the gradient difference between adjacent grids (e.g., the gradient difference between grid A and grid B is 3 people / minute) based on the flow interval of the grid cell (e.g., high flow interval, low flow interval).

[0135] The extracted rate of change features are weighted by applying optimized weight parameters (e.g., 0.65 for "Privacy Glass" and 0.48 for "Encrypted Storage"). For grids associated with high-weight privacy labels (e.g., grids related to "Privacy Glass"), the weight of their rate of change features is increased (e.g., multiplied by 1.2); for low-weight or irrelevant grids, the weight of their rate of change features is decreased (e.g., multiplied by 0.8).

[0136] After weighting, the features of time and space dimensions are integrated into a multi-dimensional vector to form a grid-adjusted feature quantity (including weighted flow change features of each time period and each interval).

[0137] Step 52 above predetermines a feature space comprising three dimensions: timeline (aligned with the live broadcast period), event type (e.g., social hotspots, public opinion events), and impact intensity (0-100 points). The event impact feature values ​​are then mapped according to these three dimensions:

[0138] Time dimension: Match the time of the event to the corresponding live broadcast time (e.g., "if a privacy and security event occurs in the second segment of the live broadcast, then it is mapped to the second segment dimension");

[0139] Event type dimension: Classified by the degree of relevance between the event and vehicle model attributes (e.g., "privacy-related events" are classified as category A, and "price-related events" are classified as category B);

[0140] Influence strength dimension: Standardize the feature value to 0-100 (e.g., if the original feature value is 80, it will still be 80 after mapping).

[0141] The mapped event impact feature values ​​are fused with grid adjustment feature values ​​in multiple dimensions:

[0142] Time dimension alignment: Ensure that the two match within the same live broadcast period (e.g., the event features of the second period are fused with the grid features of the second period);

[0143] Feature Dimension Merging: Merge the event impact intensity and the weighted rate of change of the grid adjustment feature into a tensor (e.g., “second time period + type A event + grid weighted gradient” to form a three-dimensional feature);

[0144] Integrate into a comprehensive correction parameter: Compress the dimensions of the fused tensor to generate a comprehensive correction parameter that includes the impact of events and changes in grid flow (e.g., "the comprehensive correction coefficient for the second time period is 1.15").

[0145] Step 53 above identifies the fluctuation range (such as peak range, trough range, and stable range) of the basic flow forecast curve, and adjusts the curve shape according to the comprehensive correction parameters:

[0146] For peak intervals: If the comprehensive correction parameter is positive (indicating that the event and grid features jointly drive the flow increase), then the peak height is increased according to the parameter value (e.g., parameter 1.15 corresponds to a 15% increase in peak height);

[0147] For trough intervals: If the parameter is negative (indicating the presence of factors inhibiting flow), the trough depth is reduced according to the parameter value (e.g., a parameter of 0.9 corresponds to a 10% reduction in trough depth);

[0148] For the stable interval: adjust the slope of the curve according to the parameters (e.g., a 5% increase in the slope corresponding to parameter 1.05 makes the curve more closely match the actual trend).

[0149] The adjustment results from all intervals are integrated to form a revised flow forecast curve, which reflects the combined effect of the event's impact and grid flow changes.

[0150] In step 54 above, the flow rate curve is compared and corrected for two consecutive cycles at a fixed period (e.g., every 30 minutes): the difference between the flow rate value of the next cycle and that of the previous cycle is calculated. If the difference is negative and the absolute value accounts for a proportion of the flow rate of the previous cycle greater than a preset range (e.g., 10%), it is determined to be "continuous decline".

[0151] The system calls up the vehicle model sensitivity threshold (e.g., the privacy label weight threshold for high-end vehicles is 0.5). If the optimized weight parameter (e.g., the weight of "privacy glass" is 0.65) is greater than this threshold, it indicates that users are highly concerned about the privacy label. In this case, the traffic drop needs to be addressed first.

[0152] When both "traffic decrease > 10% for two consecutive cycles" and "optimized weight parameter > 0.5" are met, a pre-trigger instruction is automatically generated that includes the target user group (users interested in high-end models) and the type of promotional strategy (such as "privacy configuration exclusive gift package").

[0153] In a preferred embodiment of the present invention, step 6 above, based on the pre-generated promotion trigger instruction and the cross-domain dataset, executes the promotion trigger instruction; simultaneously, it processes the correlation between the cross-domain dataset and the public opinion of the live stream comments using a mirroring algorithm, and executes the wording optimization instruction when the correlation exceeds a preset threshold, which may include:

[0154] Step 61: Analyze the target user group and promotion strategy type in the pre-generated promotion trigger instruction; trigger pop-up push and limited-time gift pack distribution operations through the live streaming platform interface to obtain the promotion execution status marker;

[0155] Step 62: When the promotion execution status flag detects that the promotion has been executed, activate the mirroring process of the cross-domain dataset and the real-time comment stream; based on the cross-domain dataset, construct a mapping space between public opinion keywords and comments through the mirroring algorithm; calculate the real-time correlation value within the mapping space;

[0156] Step 63: When the real-time relevance value exceeds the preset dynamic threshold, locate highly relevant public opinion keywords, generate a set of speech optimization instructions, including: keyword reinforcement explanation markers and parameter comparison speech templates; execute the speech optimization instruction set through the anchor teleprompter to complete the speech optimization closed loop.

[0157] In this embodiment of the invention, by accurately parsing instructions and targeting platform operations, it is ensured that promotional information only reaches the target user group, reducing invalid pushes; real-time acquisition of execution status provides an execution benchmark for subsequent public opinion analysis, improving the accuracy and controllability of promotions; the mapping space constructed by the mirror algorithm realizes the correlation between cross-domain public opinion and live room comments, and real-time correlation calculation can promptly capture the focus of user feedback on promotional activities; based on highly correlated keywords, targeted verbal instructions are generated, and the teleprompter drives the anchor to dynamically adjust the content of the explanation, quickly respond to user questions, improve user interaction experience and promotional conversion rate.

[0158] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0159] Step 61 above extracts the core information from the pre-generated promotion trigger instruction:

[0160] Target user group: Filter the corresponding user ID list by user tags contained in the command (such as "users who are interested in high-end models" or "users who have queried privacy configurations ≥ 5 times");

[0161] Promotional strategy type: Identify the strategy type in the instruction (such as "Exclusive Limited-Time Gift Package for Privacy Configuration" or "Instant Discount for Test Drive Appointment"), and specify the promotional content (configuration included in the gift package, discount amount, and validity period).

[0162] Call the pop-up push interface of the live streaming platform to push promotional information to the target user group ID (e.g., the pop-up content is "Exclusive for high-end users: Privacy glass upgrade gift pack available for 30 minutes only"); call the platform's limited-time gift pack distribution interface to set the gift pack distribution rules (e.g., one gift pack per person, bound to user account), and monitor the distribution status in real time (e.g., "Pushed", "Received", "Distribution failed"); integrate the push and distribution status information to generate a promotion execution status marker (e.g., "Successfully pushed to 80% of target users, 20% of users have received the gift pack").

[0163] In step 62 above, when the promotion execution status flag shows "Promotion executed" (such as pop-up push completed, gift pack distribution started), the cross-domain dataset and the real-time comment stream in the live broadcast room are automatically started for correlation analysis, and the processing cycle is set (such as once every 5 minutes).

[0164] Extract public opinion keywords (such as "price-performance ratio of privacy package" and "test drive discount") from cross-domain datasets (such as vertical automotive forum data) as "source keywords" in the mapping space; collect real-time comment streams from live broadcast rooms and extract keywords from user comments (such as "what does the package include" and "can discounts be stacked") as "target keywords" in the mapping space; construct the correspondence between source keywords and target keywords based on semantic similarity (such as "price-performance ratio" and "discount" being semantically related), forming a mapping space (such as the association network of "source keyword A - target keyword A1 / A2").

[0165] Statistical correlation strength within the mapping space:

[0166] Calculate the frequency of co-occurrence of source keywords and target keywords (e.g., "privacy package" appears 100 times in cross-domain public opinion, and related keywords appear 80 times in live broadcast comments); combine the semantic matching degree (e.g., 1 point for complete match, 0.6 points for partial match), and calculate the weighted real-time relevance value (e.g., co-occurrence frequency × semantic matching degree, the final score is 75 points).

[0167] In step 63 above, when the real-time relevance value exceeds the preset dynamic threshold (e.g., 60 points), the keywords with the highest relevance in the mapping space (e.g., "price-performance ratio of privacy package" and "test drive discount restrictions") are selected and marked as high-relevance public opinion keywords that require a key response.

[0168] For highly relevant keywords, generate tags that the anchor needs to emphasize (e.g., "The 3 upgrade configurations included in the privacy package need to be explained in detail within 1 minute"); Parameter comparison script template: Combine user questions in cross-domain public opinion (e.g., "Comparison with competitor packages") to generate structured script templates (e.g., "Our privacy package includes 1 more soundproofing upgrade than competitors, and the discount is 20% higher").

[0169] The script optimization instruction set is synchronized to the anchor's teleprompter in real time (e.g., the teleprompter displays "Focus on explaining: cost-effectiveness of the privacy package, refer to template 3"), the anchor's execution is monitored (e.g. whether the explanation is given within the specified time), and the optimization effect is confirmed based on subsequent comment feedback (e.g., a reduction in user questions), thus completing the closed loop from "public opinion correlation analysis" to "script adjustment".

[0170] like Figure 2 As shown, embodiments of the present invention also provide a live-streaming car marketing traffic fluctuation prediction system, comprising:

[0171] The acquisition module is used to simultaneously acquire data within the live streaming platform and cross-domain extended data, generating platform datasets and cross-domain datasets;

[0172] The association module is used to construct an association space between social hot topics and vehicle attributes based on cross-domain datasets and through mirroring algorithms. When the intensity of event features in the association space exceeds a preset threshold, event impact feature values ​​are generated.

[0173] The convex hull algorithm module is used to generate basic traffic prediction curves and select key point sets at fixed intervals based on event impact feature values ​​and platform datasets; and to generate convex polygon regions by processing the key point sets through the convex hull algorithm.

[0174] The optimized weight module is used to simultaneously calculate the gradient distribution matrix of the traffic change gradient of each grid cell when segmenting the convex polygon region into grid cells, and to identify the vehicle type; if it is a high-end vehicle, it calls the mirror algorithm based on the historical feature library, establishes a privacy label weight mapping relationship based on the traffic change gradient, and generates optimized weight parameters.

[0175] The correction module is used to generate grid adjustment feature quantities based on optimized weight parameters and gradient distribution matrix; it integrates event impact feature values ​​and grid adjustment feature quantities to generate comprehensive correction parameters, and corrects the basic traffic prediction curve to generate a corrected traffic curve; when the corrected traffic curve shows a downward trend in traffic and the optimized weight parameters meet the preset threshold, a promotion trigger instruction is pre-generated.

[0176] The optimization instruction module is used to execute promotion trigger instructions based on pre-generated promotion trigger instructions and cross-domain datasets; at the same time, it uses a mirroring algorithm to process the correlation between cross-domain datasets and live stream comments, and executes the wording optimization instruction when the correlation exceeds a preset threshold.

[0177] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting traffic fluctuations in live-stream car marketing, characterized in that, The method includes: Step 1: Synchronously collect data within the live streaming platform and cross-domain extended data to generate platform datasets and cross-domain datasets; Step 2: Based on the cross-domain dataset, construct the association space between social hot topics and vehicle attributes using the mirroring algorithm. When the event feature strength in the association space exceeds a preset threshold, generate event impact feature values. Step 3: Based on the event impact feature values ​​and platform dataset, generate basic traffic prediction curves at fixed intervals and select key point sets; process the key point sets using the convex hull algorithm to generate convex polygon regions; Step 4: When segmenting the convex polygon region into grid cells, simultaneously calculate the traffic change gradient of each grid cell to generate a gradient distribution matrix and identify the vehicle type; if it is a high-end vehicle, call the mirror algorithm based on the historical feature library, establish a privacy label weight mapping relationship based on the traffic change gradient, and generate optimized weight parameters. Step 5: Generate grid adjustment feature quantity based on optimized weight parameters and gradient distribution matrix; merge event impact feature value and grid adjustment feature quantity to generate comprehensive correction parameter, and correct the basic traffic prediction curve to generate corrected traffic curve; when the corrected traffic curve shows a downward trend in traffic and the optimized weight parameter meets the preset threshold, pre-generate promotion trigger instruction; Step 6: Execute the promotion trigger instruction based on the pre-generated promotion trigger instruction and cross-domain dataset; at the same time, process the public opinion correlation between the cross-domain dataset and the live broadcast room comments through the mirroring algorithm, and execute the wording optimization instruction when the correlation exceeds the preset threshold.

2. The method for predicting traffic fluctuations in live-stream car marketing according to claim 1, characterized in that, Step 1: Synchronously collect data from the live streaming platform and cross-domain extended data to generate platform datasets and cross-domain datasets, including: Step 11: Collect platform dynamic data in real time from the pre-live broadcast period to the end of the live broadcast to obtain raw platform data; Step 12: Based on the original platform data, simultaneously trigger cross-domain data collection of the target model's public opinion keyword set in the vertical automotive forum and the consultation volume trend data of offline sales channels to obtain the original cross-domain data. Step 13: Based on the original cross-domain data and the original platform data, align the original platform data according to the live broadcast timeline to generate time-aligned platform data; based on the time-aligned platform data, extract the interaction rate-conversion rate correlation matrix; synchronize the original cross-domain data with the aligned platform timeline to generate time-aligned cross-domain data. Step 14: Perform feature encoding on time-aligned platform data to generate a platform dataset containing a peak viewing distribution matrix, an interaction conversion correlation tensor, and a competitor traffic time-series feature vector; perform semantic aggregation on time-aligned cross-domain data to generate a cross-domain dataset containing public opinion keyword weight distribution data and an offline consultation volume gradient sequence.

3. The method for predicting traffic fluctuations in live-stream car marketing according to claim 2, characterized in that, Step 2: Based on the cross-domain dataset, construct an association space between social hot topics and vehicle attributes using a mirroring algorithm. When the event feature strength in the association space exceeds a preset threshold, generate event impact feature values, including: Step 21: Based on the cross-domain dataset, analyze the public opinion keyword weight distribution data and extract the high-frequency keyword set; standardize the offline consultation volume gradient sequence and generate a consultation volume change vector to obtain the preprocessed cross-domain dataset. Step 22: Based on the preprocessed cross-domain dataset, the high-frequency keyword set is mapped to the vehicle attribute space using the mirroring algorithm. The consultation volume change vector is then subjected to tensor product operation with the vehicle attribute space to obtain the social hotspot-vehicle attribute association space. Step 23: Calculate the feature vector magnitude of each social hotspot event in the social hotspot-vehicle attribute association space, normalize the magnitude, and generate a sequence of event feature intensity values. Step 24: When any intensity value in the event feature intensity value sequence exceeds a preset dynamic threshold, feature value generation is triggered and the intensity value exceeding the threshold is converted into event impact feature value.

4. The method for predicting traffic fluctuations in live-stream car marketing according to claim 3, characterized in that, Step 3: Based on the event impact feature values ​​and platform dataset, generate a basic traffic prediction curve at a fixed period and select a key point set; process the key point set using the convex hull algorithm to generate a convex polygon region, including: Step 31: Inject the event impact feature values ​​into the interaction transformation association tensor of the platform dataset, and generate a spatiotemporal augmented dataset based on the fused tensor; Step 32: Based on the spatiotemporal augmentation dataset, perform time series prediction at fixed intervals to generate a basic traffic prediction curve; correct the shape characteristics of the basic traffic prediction curve based on the historical high-end vehicle feature library, and automatically select a set of key points, where: peak points correspond to promotional sensitive periods, trough points correspond to periods of traffic loss risk, and inflection points correspond to periods of public opinion change response. Step 33: Select boundary extreme points in the key point set using the convex hull algorithm, and connect them in chronological order to generate an initial convex polygon region; during the connection process, add auxiliary connection points to the key point set corresponding to the privacy label period of high-end models to improve the region accuracy, and simplify the connection path to optimize the calculation efficiency for the key point set corresponding to the price sensitive period of economy models, thereby generating a convex polygon region covering the core traffic fluctuation area.

5. The method for predicting traffic fluctuations in live-stream car marketing according to claim 4, characterized in that, Step 4: When dividing the convex polygon region into grid cells, simultaneously calculate the gradient distribution matrix of the flow change gradient of each grid cell and identify the vehicle type; For high-end models, the mirroring algorithm is invoked based on the historical feature database to establish a privacy label weight mapping relationship according to the traffic change gradient, generating optimized weight parameters, including: Step 41: Based on the convex polygon region, divide it into grid cells along the time axis, and extract the flow data and time period information contained in each grid cell to generate grid cell data; simultaneously calculate the flow change rate of each grid cell to generate a gradient distribution matrix; Step 42: Based on the grid cell data and gradient distribution matrix, extract user profile features within the grid cells, including user query behavior and interaction frequency regarding vehicle configuration; call the historical high-end vehicle feature library for feature comparison, count the frequency of privacy configuration queries, and when the frequency is greater than the benchmark value in the historical high-end vehicle feature library, mark it as a high-end vehicle; otherwise, mark it as a Volkswagen vehicle to obtain the vehicle identification. Step 43: When the vehicle model is identified as a high-end vehicle, the gradient distribution matrix is ​​invoked and a privacy label weight mapping relationship is established based on the historical high-end vehicle feature library through the mirroring algorithm. That is, the gradient change rate is used as the independent variable, the historical weight of the privacy label in the feature library is used as the base value, and the weight value is dynamically adjusted according to the gradient change direction. When the gradient increases, the weight increase is increased, and when the gradient decreases, the weight decrease is applied to generate optimized weight parameters.

6. The method for predicting traffic fluctuations in live-stream car marketing according to claim 5, characterized in that, Step 5: Generate grid adjustment feature quantities based on optimized weight parameters and gradient distribution matrix; fuse event impact feature values ​​and grid adjustment feature quantities to generate comprehensive correction parameters, and correct the basic flow prediction curve to generate the corrected flow curve; When the traffic curve shows a downward trend and the optimized weight parameters meet the preset threshold, a promotion trigger instruction is pre-generated, including: Step 51: Based on the optimized weight parameters and gradient distribution matrix, extract the rate of change feature through the spatiotemporal characteristics of the gradient distribution matrix, and perform weighted aggregation on the rate of change feature based on the optimized weight parameters to obtain the grid adjustment feature quantity. Step 52: Based on the mesh adjustment feature quantity and the event impact feature value, map the event impact feature value to the preset feature space, and perform tensor fusion with the mesh adjustment feature quantity to obtain the comprehensive correction parameter; Step 53: Inject the comprehensive correction parameters into the fluctuation range of the basic flow prediction curve to generate the corrected flow curve; Step 54: When it is detected that the flow rate decreases for two consecutive cycles in the corrected flow rate curve is greater than the preset range, and the verification optimization weight parameter is greater than the vehicle model sensitivity threshold, a promotion trigger instruction is pre-generated.

7. The method for predicting traffic fluctuations in live-stream car marketing according to claim 6, characterized in that, Step 6: Based on the pre-generated promotion trigger command and cross-domain dataset, execute the promotion trigger command; simultaneously, process the correlation between the cross-domain dataset and the live stream comments using a mirroring algorithm. When the correlation exceeds a preset threshold, execute the script optimization command, including: Step 61: Analyze the target user group and promotion strategy type in the pre-generated promotion trigger instruction; trigger pop-up push and limited-time gift pack distribution operations through the live streaming platform interface to obtain the promotion execution status marker; Step 62: When the promotion execution status flag detects that the promotion has been executed, activate the mirroring process of the cross-domain dataset and the real-time comment stream; based on the cross-domain dataset, construct a mapping space between public opinion keywords and comments through the mirroring algorithm; calculate the real-time correlation value within the mapping space; Step 63: When the real-time relevance value exceeds the preset dynamic threshold, locate highly relevant public opinion keywords, generate a set of speech optimization instructions, including: keyword reinforcement explanation markers and parameter comparison speech templates; execute the speech optimization instruction set through the anchor teleprompter to complete the speech optimization closed loop.

8. A live-streaming car marketing traffic fluctuation prediction system, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to simultaneously acquire data within the live streaming platform and cross-domain extended data, generating platform datasets and cross-domain datasets; The association module is used to construct an association space between social hot topics and vehicle attributes based on cross-domain datasets and through mirroring algorithms. When the intensity of event features in the association space exceeds a preset threshold, event impact feature values ​​are generated. The convex hull algorithm module is used to generate basic traffic prediction curves and select key point sets at fixed intervals based on event impact feature values ​​and platform datasets; and to generate convex polygon regions by processing the key point sets through the convex hull algorithm. The weighting module is optimized to simultaneously calculate the gradient distribution matrix of the traffic flow change gradient of each grid cell when the convex polygon region is segmented into grid cells, and to identify the vehicle type. For high-end models, the mirror algorithm is called based on the historical feature library to establish a privacy label weight mapping relationship according to the traffic change gradient and generate optimized weight parameters. The correction module is used to generate grid adjustment features based on optimized weight parameters and gradient distribution matrix; The event impact feature value and grid adjustment feature value are merged to generate a comprehensive correction parameter, and the basic traffic prediction curve is corrected to generate a corrected traffic curve; when the corrected traffic curve shows a downward trend in traffic and the optimized weight parameter meets the preset threshold, a promotion trigger instruction is pre-generated. The optimization instruction module is used to execute promotion trigger instructions based on pre-generated promotion trigger instructions and cross-domain datasets; at the same time, it uses a mirroring algorithm to process the correlation between cross-domain datasets and live stream comments, and executes the wording optimization instruction when the correlation exceeds a preset threshold.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images