A live broadcast e-commerce product advertisement pushing method

Abstract

The present application relates to the field of advertisement pushing, and particularly relates to a live broadcast e-commerce product advertisement pushing method. The method comprises the following steps: collecting a viewing behavior data stream; performing feature recognition based on the viewing behavior data stream to generate viewing state description data; performing interest preference analysis based on the viewing state description data to obtain user preference content data; performing commodity matching degree calculation matching according to the user preference content data to generate a commodity pushing sequence; performing advertisement putting time analysis based on the viewing behavior data stream to output an advertisement putting interval; and embedding and pushing the commodity pushing sequence according to the advertisement putting interval. The present application realizes accurate matching of commodity content and advertisement putting time, improves advertisement benefit conversion rate and user live broadcast viewing experience.

Description

Technical Field

[0001] This invention relates to the field of advertising push, and more particularly to a method for pushing advertisements for live e-commerce products. Background Technology

[0002] In actual live streaming, user viewing behavior is highly dynamic and personalized, with significant differences among users in viewing duration, interaction frequency, content focus, and purchase intent. If advertising or product recommendations lack targeting, issues such as mismatched content with user interests and inappropriate timing can easily arise, leading to a decline in user experience, poor advertising effectiveness, and even user resentment, impacting the platform's overall operational performance. Existing live e-commerce advertising methods often rely on simple user tags or historical purchase records for recommendations, frequently ignoring real-time viewing behavior and changes in current interests during the live stream. Furthermore, some advertising strategies use fixed time points or manual rules for delivery, lacking analysis of the live stream's pace and changes in user attention, making it difficult to effectively integrate advertising content with the live stream content, resulting in generally low advertising conversion rates and delivery efficiency. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a method for pushing product advertisements in live-streaming e-commerce, thereby resolving at least one of the aforementioned technical issues.

[0004] To achieve the above objectives, the present invention provides a method for pushing product advertisements in live-streaming e-commerce, comprising the following steps: Step S1: Collect viewing behavior data stream; perform feature recognition based on viewing behavior data stream to generate viewing status description data; Step S2: Perform interest and preference analysis based on viewing status description data to obtain user preferred content data; Step S3: Calculate and match product suitability based on user preference content data to generate a product push sequence; Step S4: Analyze ad delivery time based on viewing behavior data stream and output ad delivery intervals; Step S5: Embed the product push sequence according to the advertising delivery interval.

[0005] The beneficial effects of this invention are specifically as follows: By collecting viewing behavior data streams (such as dwell time, swipe frequency, likes, comments, and replays), it can objectively reflect the user's actual participation and attention status during the live stream. Feature identification of viewing behavior data and generation of viewing status description data helps transform discrete and complex behavioral data into structured information, providing a foundation for subsequent analysis and decision-making. Analysis of viewing status description data can accurately determine user preferences for different content types, product categories, or streamer styles, improving the targeting of interest identification. Preference content data based on users' actual viewing behavior helps reduce product pushes unrelated to user interests, improving the effectiveness of advertising. By calculating the fit between products and user-preferred content, products that better meet user needs can be selected, improving the relevance of push notifications. Generating a product push sequence allows for sorting products based on fit, prioritizing products more likely to trigger purchase behavior. A precisely matched product push sequence helps improve user click-through rates and purchase conversion rates, enhancing the commercial value of live-streaming e-commerce. By analyzing viewing behavior data streams, we can determine the time intervals when users are most focused on watching and most actively interacting, avoiding ad placement during periods of low user attention. Pushing ads within appropriate timeframes helps improve ad visibility and user acceptance, reducing the likelihood of ads being ignored or skipped. Embedding product push sequences into reasonable ad placement intervals makes ad display smoother and avoids abrupt interruptions to the live stream content. Pushing products during times of high user attention and matching interests helps enhance user identification and trust in the ad content. Through dual optimization of time and content matching, we can achieve precise ad targeting and efficient conversion, improving the overall operational effectiveness of live-streaming e-commerce. Attached Figure Description

[0006] Figure 1 This is a flowchart illustrating the steps of a live-stream e-commerce product advertising push method according to the present invention. Figure 2 This is a detailed flowchart illustrating the implementation steps of step S1. Figure 3 This is a flowchart illustrating the detailed implementation steps of step S2. Detailed Implementation

[0007] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0008] This application provides a method for pushing product advertisements in live-streaming e-commerce. The execution entities of this method include, but are not limited to, mechanical equipment, data processing platforms, cloud server nodes, and network upload devices that can be considered general computing nodes in this application. The data processing platform includes, but is not limited to, at least one of an audio / image management system, an information management system, and a cloud-based data management system.

[0009] Please see Figures 1 to 3 This invention provides a method for pushing product advertisements in live-streaming e-commerce, comprising the following steps: Step S1: Collect viewing behavior data stream; perform feature recognition based on viewing behavior data stream to generate viewing status description data; Step S2: Perform interest and preference analysis based on viewing status description data to obtain user preferred content data; Step S3: Calculate and match product suitability based on user preference content data to generate a product push sequence; Step S4: Analyze ad delivery time based on viewing behavior data stream and output ad delivery intervals; Step S5: Embed the product push sequence according to the advertising delivery interval.

[0010] In the embodiments of the present invention, see Figure 1 The diagram below illustrates the steps of a live-streaming e-commerce product advertising push method according to the present invention. In this example, the steps of the live-streaming e-commerce product advertising push method include: Step S1: Collect viewing behavior data stream; perform feature recognition based on viewing behavior data stream to generate viewing status description data; In this embodiment, user viewing behavior is continuously collected through a live streaming platform to form a real-time viewing behavior data stream. The data stream includes the timestamp of the user entering the live stream, changes in video playback status, dwell time, bullet screen commenting, likes, add-to-cart, and follow interactions, as well as information such as video frame switching frequency and changes in the host's speaking speed. The data stream is cleaned and standardized, for example, by standardizing dwell time to the second level, encoding interactive events by type, and vectorizing bullet screen semantics. Behavior recognition is performed based on multi-dimensional features, including calculating the user's viewing stability, attention duration, and interaction response rate within a different window (e.g., a sliding window every 5 minutes with 50% overlap). Attention duration can be calculated by combining dwell time and interaction frequency, while the interaction response rate is the number of likes, add-to-cart, and bullet screen comments per unit time. A viewing status description data vector is generated, with each record containing the user ID, time period, dwell time, interaction metrics, and attention characteristics.

[0011] Step S2: Perform interest and preference analysis based on viewing status description data to obtain user preferred content data; In this embodiment, user dwell time, interaction frequency, and attention metrics across different live streams and content tags are normalized to eliminate differences in the dimensions of these metrics. A multi-dimensional weighted scoring method is used to calculate user preference values ​​for each content category, for example, dwell time is weighted at 0.5, interaction frequency at 0.3, and attention duration at 0.2. Preference values ​​are accumulated across different content tags to form a user preference matrix. Furthermore, clustering or ranking methods can be used to identify concentrated areas of user interest, such as user groups with strong preferences for entertainment, digital products, or clothing. User preference content data, including content tags, preference intensity, and interaction activity index, is generated to provide quantitative decision-making basis for product adaptation and advertising placement.

[0012] Step S3: Calculate and match product suitability based on user preference content data to generate a product push sequence; In this embodiment, product fit calculation is performed by matching user preference content data with product database features. Product attributes include category, price range, functional characteristics, and display scenarios. After vectorizing product attributes, similarity is calculated between them and user preference vectors. Common methods include cosine similarity or weighted matching, where user interest weights are allocated based on preference intensity, and product attribute weights are set based on display focus. For example, for tags with preference intensity ≥ 0.7, the weight of that product category can be set to 0.6, and the weight of other product categories can be 0.4. The matching results are sorted from high to low fit to generate a product push sequence. Each product record in the sequence includes product ID, push priority, estimated click-through rate, and fit score. During the push sequence generation process, diversity constraints can be introduced, such as avoiding the continuous display of products of the same category or similar price range, to enhance user viewing experience and conversion potential.

[0013] Step S4: Analyze ad delivery time based on viewing behavior data stream and output ad delivery intervals; In this embodiment, ad delivery time analysis is performed based on user viewing behavior patterns and attention distribution. User viewing start time, dwell time, and peak attention periods are extracted from viewing status description data. For example, each peak period is defined as a range where continuous viewing lasts ≥3 minutes and the interaction frequency is 1.5 times higher than the average. The ad delivery interval is set during peak user attention periods, and a sliding time window method (window length 1–2 minutes, 50% overlap) can be used to detect potentially high-acceptance intervals. The ad acceptability probability is calculated for each window, and a peak detection algorithm is used to identify peak acceptance points. The ad delivery interval is output, including start time, end time, and estimated acceptance, providing a precise time range for product push embedding. Furthermore, the interval length can be dynamically optimized based on user tolerance cycles, viewing duration, and interaction behavior. For example, it can be extended by 1–3 minutes for highly active users and shortened for less active users to avoid fatigue, ensuring that ads are delivered when user attention is highest.

[0014] Step S5: Embed the product push sequence according to the advertising delivery interval.

[0015] In this embodiment, the product push sequence is embedded within the advertising period, achieving personalized pushes through time management and sequence control. The total length of the advertising period is calculated, and the push duration for each product is set to 5–15 seconds. The number of products that can be pushed is calculated based on the period length and the individual product duration. For example, if the advertising period is 40 seconds and each product is displayed for 8 seconds, then 5 products can be pushed. A product list is generated by extracting the corresponding number of products according to the priority order of the product push sequence. A corresponding advertising video is generated for each product, including product images, prices, functional descriptions, and display scenarios. This can be batch-produced using templated video generation methods. The generated videos are embedded into the live stream content and played using either interstitial or overlay recommendation slots to ensure that the advertising push is synchronized with the user's viewing rhythm. User feedback data on the pushed videos is collected in real time for subsequent advertising effectiveness evaluation and dynamic adjustments, achieving a personalized, precise, and quantifiable product push strategy.

[0016] In this embodiment, see Figure 2 The diagram below illustrates the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include: The live streaming platform monitors the user's viewing process in real time and collects viewing behavior data streams, including the distribution of dwell time, changes in the semantic rhythm of bullet comments, the density of interactive operations, and the frequency of screen switching. Perform time-series serialization processing on the viewing behavior data stream to construct the viewing behavior trajectory; Based on the viewing behavior trajectory, the viewing stability, interaction response rate, and attention duration are calculated to generate viewing state description data.

[0017] In this embodiment, viewing behavior information is continuously collected through front-end tracking and back-end data interfaces, including dwell time, bullet comment sending, likes, add-to-cart, and follow records, tracking screen switching frequency and playback status. Dwell time is recorded by accumulating the viewing time of each video segment precisely in milliseconds; bullet comment semantic rhythm is annotated by the sending time interval, length, and lexical sentiment of each bullet comment and mapped to numerical indicators to quantify user attention fluctuations; interaction operation density is calculated by counting the number of likes, add-to-cart, and follow actions per unit time to reflect user participation; screen switching frequency is calculated by recording the timestamps of video frame changes or scene switching events, serving as a reference for attention concentration and information processing speed. Behavioral data is collected in time series form, with each time point containing multi-dimensional behavioral characteristics, enabling continuous tracking of viewing behavior. Dwell time, bullet comments, likes, add-to-cart, and follow actions are categorized using a 500-millisecond sliding window, with a window overlap rate set to 50% to ensure the continuity of short-term behavioral changes. The bullet screen text content undergoes word segmentation, sentiment analysis, and semantic rhythm annotation to transform textual information into numerical features, generating rhythm indicators. Interaction density is summarized per unit time to obtain the operation frequency within each window; the number of screen transitions for each window is calculated, forming a frequency sequence. Outliers are removed by calculating the standard deviation of behavior within each time window; for example, continuous abnormally long periods of time spent on screen or repeated operations are excluded. Through multi-dimensional time series integration, the user's behavioral state throughout the entire live stream viewing process is continuously represented as a behavioral trajectory. The trajectory at each time point includes multi-dimensional information such as time spent on screen, bullet screen rhythm, interaction frequency, and screen transitions.

[0018] Core metrics for user viewing are extracted from behavioral trajectories. Viewing stability is calculated using the fluctuation coefficient or entropy value of dwell time; a smaller fluctuation coefficient indicates a smoother viewing process. Interaction response rate is calculated by statistically analyzing the intervals between actions such as liking, commenting, and adding to cart within a continuous time window, yielding the average number of responses per unit time. Multi-scale analysis using three window lengths (1 second, 3 seconds, and 5 seconds) is conducted to reflect instantaneous and short-term reaction capabilities. Attention duration is judged by the level of behavioral activity within a continuous time window. When dwell time exceeds a threshold, comment input is frequent, or interactive operations are intensive, attention is considered to be at a high level; the cumulative time of the continuous window is the attention duration. Viewing stability, interaction response rate, attention duration, and the semantic rhythm features of comments are integrated to form complete viewing state description data.

[0019] In this embodiment, see Figure 3 The diagram below illustrates the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include: Identify multiple live streaming rooms based on viewing behavior data streams; The live stream content of the multiple live streaming rooms is analyzed to obtain multiple content tags; Calculate user activity levels in different live streaming rooms based on viewing behavior data streams; Based on the behavioral activity level, multiple content tags are associated and labeled to obtain the interaction intensity coefficient of different live streaming rooms; Based on the viewing status description data and the interaction intensity coefficient, interest preference analysis is performed to obtain user preferred content data.

[0020] In this embodiment, the timestamps of users entering and leaving the live stream are continuously tracked. Viewing segments that stay for more than a certain threshold (e.g., 10 seconds) are marked as valid visits, and different live streams are distinguished according to the order of access time. The behavior of continuously accessing the same live stream is aggregated, integrating information such as dwell time, interactive behavior, and bullet screen input to form a user behavior record for that live stream. Furthermore, the type and frequency of operation events in the behavior flow can also help determine whether a user is truly participating in watching or only browsing briefly, thereby excluding low-activity or accidental touch behaviors. For cases where a user accesses multiple live streams within the same time period, a time segmentation strategy is adopted, recording the live streams corresponding to each time period separately and generating a list of multiple live streams, forming a cross-live stream behavior matrix for each user. For each live stream, multi-dimensional content analysis is performed using video content, host voice, and text information. Video footage is analyzed using keyframe extraction and visual feature recognition technologies, such as identifying product categories, scene types, and activity elements. A key frame is extracted every 1-2 seconds, and visual tags are generated using image recognition algorithms. The broadcaster's audio content undergoes speech-to-text extraction to extract semantics, combined with natural language processing for keyword extraction, intent recognition, and sentiment analysis, generating audio tags. Bullet comments and interactive text information also undergo word segmentation and semantic analysis, mapping high-frequency keywords, sentiment trends, and interactive topics to tags. All tags are integrated using a multimodal fusion method to form a content tag set for each live stream, including attributes such as product categories, activity themes, and interaction formats, providing accurate data for subsequent analysis of user behavior and content correlation. During tag generation, frequency thresholds can be set (e.g., keywords appearing more than 3 times or accounting for more than 5% of total bullet comments) to filter out low-relevant information, ensuring the representativeness and stability of the tags.

[0021] Behavioral activity is quantified using metrics such as user dwell time in each live stream, interaction density, and bullet screen comment frequency. Dwell time is calculated by dividing total viewing time by the total live stream duration, reflecting the user's level of engagement with the content. Interaction density is calculated as the average frequency of likes, purchases, and follows per unit time. Bullet screen comment frequency is quantified by the number of bullet screen comments sent per minute, weighted by bullet screen comment length and semantic activity. The comprehensive behavioral activity index uses a normalization method, mapping dwell time, interaction density, and bullet screen comment activity to a 0-1 scale, and then summing them according to their weights. For example, dwell time accounts for 40%, interaction density for 35%, and bullet screen comment activity for 25%. This method allows us to obtain the behavioral activity value for each user in each live stream, distinguishing between high-engagement and low-engagement live streams. By mapping user behavioral activity in each live stream to live stream content tags, an interaction intensity coefficient is generated. The interaction intensity corresponding to each tag is calculated using a weighted average. This involves summing the user's activity levels in live streams containing that tag, normalizing the sum to obtain a coefficient between 0 and 1, representing the user's interest intensity in that content tag. For example, if a user's activity levels in three live streams containing the tag "beauty and cosmetics" are 0.8, 0.6, and 0.9 respectively, then the weighted average interaction intensity for that tag can be calculated as 0.77. For live streams with multiple tags, a multi-tag weighting strategy is used, assigning different weights to each tag based on its frequency or importance, such as a weight of 0.6 for high-frequency tags and 0.4 for low-frequency tags.

[0022] Interest preference analysis comprehensively evaluates users' viewing status data (including viewing stability, interaction response rate, and attention duration) and content tag interaction intensity coefficients. Viewing status data reflects users' focus and engagement in the live stream, and this data is used as weight to adjust the interaction intensity coefficients. For example, live stream behaviors with high stability and fast response rates can increase the preference weight of the corresponding tag. By weighting and summing each tag, a preference score for each tag is calculated, and the preference scores are normalized to 0–1 to form a user preference content data vector.

[0023] In this embodiment, step S3 includes the following steps: Extract the product database; parse the product attributes of each product in the product database and output the attribute information of all products; The attribute information includes product category, price range, functional characteristics, and display scenarios; Based on user preference content data, the attribute information is matched and multiple suitable products are extracted. Predict purchase intentions for multiple compatible products and generate potential purchase scores for different products; The product recommendation sequence is generated by performing gradient sorting based on potential purchase scores.

[0024] In this embodiment, a complete access to and information extraction of the e-commerce platform's product database is performed. The data for each product includes product ID, category level, price information, functional description, sales history, and display scenario information. Data collection is conducted via API or batch export, unifying structured information (such as category and price) and unstructured information (such as functional description, display scenario text, or images). For functional characteristics and display scenarios, key features can be extracted using natural language processing and image recognition technologies. For example, keywords, phrases, and product selling point tags can be extracted from product description text; scene types (such as kitchen, office, outdoor) and color styles can be identified from product display images to form scene tags. After extraction, all product information is stored in a standardized structured table, with each record containing multi-dimensional information such as product ID, category tag, price range, functional characteristic vector, and display scenario vector. For duplicate or missing data in the product database, deduplication and missing value imputation methods are used. For example, missing prices can be filled with the average of similar categories, and missing function descriptions can be filled with text from similar products to ensure the completeness and consistency of product information. Each product undergoes attribute parsing to generate a complete set of attribute information, including category, price range, functional characteristics, and display scenario. Category parsing uses multi-level directory mapping, for example, mapping "electronic products > headphones > wireless Bluetooth headphones" to three levels of tags, facilitating matching user preferences with product categories. Price ranges divide product prices into tiers, such as 0–100 yuan, 101–500 yuan, 501–1000 yuan, etc., converting continuous prices into discrete range tags for easier preference weight calculation. Functional characteristic parsing uses word segmentation, keyword extraction, and vectorization (TF-IDF or semantic embedding) of the product description text to form a functional characteristic vector, representing the product's core selling points and usage scenarios. Display scenario parsing combines product images and description text, generating a scene vector through multimodal feature extraction, including scene type, usage environment, and visual style.

[0025] User preference content data vectors are matched with product attribute information to generate a suitability index. The matching method uses weighted similarity calculation, integrating category tag matching, price range matching, functional feature similarity, and display scenario similarity according to weights. Category tag matching calculates a tag consistency score, e.g., 1 point for a perfect match and 0.5 points for a partial match; price range matching calculates a suitability score based on the degree of price range consistency; functional feature and display scenario vectors are converted to similarity scores using cosine similarity or Euclidean distance. Weights can be set to 0.4 for category, 0.3 for functional features, 0.2 for price, and 0.1 for display scenario to comprehensively reflect the degree to which users' interests emphasize different product attributes. Purchase intention is predicted for suitable products by comprehensively analyzing user interaction history with products, preference attributes, and behavioral data. Purchase intention is generated using a multi-factor scoring model, considering factors including the frequency of user interaction with similar products (likes, adding to cart, purchase history), the matching degree between user preference content vectors and product attribute vectors, and the current live stream behavior status (viewing stability, interaction response rate, attention duration). By converting these factors into numerical indicators and then weighting and summing them, a potential purchase score for each product can be generated, ranging from 0 to 1. A higher score indicates a stronger potential purchase intention from the user. For example, the weights can be set as follows: historical behavior 0.5, preference matching degree 0.3, and viewing status 0.2, to comprehensively reflect the user's interest in the product and the likelihood of a decision.

[0026] The potential purchase rating vectors are sorted in descending order to form a product gradient ranking sequence. The sorting method can be a simple score-based sort, or a gradient-weighted method can be used to refine the differences in ratings between adjacent products, ensuring that products with high potential purchase intent are prioritized in the push sequence, while considering the balance between product category diversity and display scenarios. After sorting, a product push sequence is generated, with each product accompanied by its potential purchase rating and matching attribute information, serving as a priority list for live-stream e-commerce pushes. The sequence can be used for real-time pushes or batch recommendations; for example, the top five products can be prioritized for display in the recommendation slot or trigger a pop-up window, with the next best products as supplementary recommendations.

[0027] In this embodiment, step S4 includes the following steps: Obtain user identity information; perform identity clustering analysis based on user identity information to obtain user type characteristics; Based on the viewing behavior data stream, the viewing tolerance cycle, fatigue time point and peak activity period are calculated to obtain the user's viewing rhythm information; Based on user viewing rhythm information and user type characteristics, conduct in-depth user evaluation and construct user feature profiles; Based on user profile characteristics, predict ad placement and output ad placement ranges.

[0028] In this embodiment, complete user identity information is obtained through user registration information, login records, and third-party data interfaces. Identity information includes basic information (such as age, gender, and geographic location), socioeconomic attributes (such as occupation category and income level), device attributes (such as terminal type and operating system version), and historical behavior records (such as historical purchase preferences and browsing categories). Textual or numerical data undergoes unified encoding processing; for example, age is divided into 10-year intervals, regions are grouped and coded by city, and occupation categories are standardized with uniform labels. Device attributes are generated into vectorized representations by identifying terminal model, screen size, and resolution, used to analyze viewing habits. After vectorizing the acquired identity information, clustering analysis methods are used to classify users. Clustering methods can include K-means, hierarchical clustering, or density clustering (DBSCAN), selecting the most suitable algorithm based on the feature distribution. For example, age, gender, region, occupation, and device attributes are combined into a multi-dimensional feature vector, which is first normalized, mapping each dimension to the 0–1 interval to eliminate dimensional differences. Then, the optimal number of clusters is determined based on the silhouette coefficient or elbow method, dividing users into different groups, such as young white-collar workers, highly active students, and housewives. Each cluster center vector serves as a representative of the user type characteristics, containing typical features of the user group in terms of age, occupation, device preferences, etc.

[0029] Viewing rhythm information is generated by analyzing users' continuous viewing behavior and interaction patterns. Viewing tolerance period is determined by statistically analyzing the continuous effective dwell time of users in the live stream, calculating the average dwell time and maximum continuous viewing time to characterize users' sustained viewing ability. Fatigue time points are identified by analyzing behavioral decay indicators, such as gradually shortening dwell time, decreasing interaction density, or reduced frequency of bullet comments, to determine the time window for user viewing fatigue. Peak activity periods are obtained by segmenting daily or weekly viewing behavior into time periods, calculating the number of users staying in each time period, interaction density, and bullet comment volume, thus obtaining the peak time period distribution.

[0030] By fusing user type characteristics with viewing rhythm information, a comprehensive user profile is generated. Fusion methods include feature vector concatenation or weighted combination. Viewing rhythm information can influence user activity scores according to weights, while user type characteristics can serve as a reference for preference and spending potential scores. In-depth user evaluation is calculated using multi-dimensional indicators, such as a comprehensive weighted score of viewing stability, interaction response rate, attention span, tolerance period, fatigue time, and peak activity periods. This score, combined with user type characteristics, quantifies potential spending power and interest preferences. The weighting can be set to 0.6 for viewing behavior and 0.4 for type characteristics, ensuring that behavioral characteristics dominate the in-depth evaluation. Ad placement prediction determines the most suitable ad display time slots and priorities by analyzing activity, viewing rhythm, and interest preference data from the user profile. The method includes calculating the ad visibility window based on peak activity periods and tolerance cycles, setting ad delivery intervals during peak user attention periods, such as the first 10 minutes of peak hours or the middle of the tolerance cycle, to improve click-through rates and conversion rates; matching ad categories based on interest preference tags to ensure ad content highly aligns with users' potential needs; and adjusting ad frequency and density by incorporating user depth ratings, increasing push frequency for high-rated users and decreasing or delaying pushes for low-rated users. Ad delivery intervals are generated through multi-dimensional rules and weighted strategies, including daily or weekly start and end times, recommended ad priority, and content categories.

[0031] In this embodiment, the specific steps for predicting ad placement based on user feature profiles and outputting the optimal ad placement range are as follows: Based on user feature profiles, predict the probability of ad acceptance and generate an ad acceptance probability curve; Peak points of acceptance were marked based on the ad acceptance probability curve; Calculate the timestamp of the peak acceptance point; Determine the user's viewing start timestamp based on the viewing behavior data stream; The advertising time is calculated based on the user's viewing start timestamp and the timestamp, and the advertising placement interval is output.

[0032] In this embodiment, the ad acceptance probability prediction comprehensively evaluates the matching degree between user behavior patterns and ad content by analyzing activity scores, viewing rhythm information, and interest preference tags in the user feature profile. Similarity is calculated between ad content and user interest tags, for example, using a vectorized matching method to calculate the cosine similarity between the ad category vector and the user interest preference vector, resulting in a matching index. Then, combined with user viewing rhythm information, the activity scores within different time windows are weighted to form a time-series-based acceptance probability value. Each time window can be set to a length of 1 minute or 5 minutes, and the sliding window overlap rate can be set to 50% to smooth short-term fluctuations and capture user attention peaks. The ad acceptance probability curve consists of the probability values ​​of each time window, reflecting the changing trend of the user's potential acceptance of the ad content throughout the viewing process. On the generated ad acceptance probability curve, the time point with the highest user acceptance of the ad is identified using a peak detection method. The curve is smoothed, for example, using moving averages or Gaussian filtering to eliminate interference from short-term fluctuations; then, a local maximum detection algorithm is used to mark the acceptance probability points in the curve that are significantly higher than those of surrounding time points as peak points. A minimum peak height threshold can be set, such as an acceptance probability greater than 0.7 and a peak interval greater than 5 minutes, to ensure that the marked peaks truly reflect users' periods of high attention rather than random fluctuations. Each peak point can record its corresponding time window and probability value, forming a peak time list.

[0033] Based on the marked peak points, time windows are mapped to actual timestamps to form precise times when users can accept the ads. The timestamp of each peak point is calculated by adding the peak's position offset within the window to the start time of the window containing the peak. For example, if the window length is 5 minutes and the peak is in the middle of the window, the peak timestamp is the window start time plus 2 minutes and 30 seconds. For the same user, there may be multiple peak points. The peak with the highest acceptance probability is selected as the primary delivery time by sorting, or a multi-point delivery strategy can be formed by combining the second highest peak. Precise timestamp calculation ensures that ad delivery is synchronized with peak user attention, avoiding delivery during periods of user fatigue or low activity, thereby improving ad display effectiveness and click-through rates. The user's viewing start time is determined by analyzing the live stream entry records, video playback events, and the time of the first interaction in the viewing behavior data stream. Specific methods include identifying the earliest timestamp of a user entering the live stream, the time when the video playback status changes, and the time when the user first likes or sends a comment, taking the earliest of these as the viewing start time. Short-term absences or accidental clicks are excluded; for example, browsing sessions lasting less than 10 seconds are not counted towards the official start time, ensuring that the start time accurately reflects the user's effective viewing behavior. The viewing start timestamp provides a benchmark for calculating ad delivery time and can be combined with the peak receptiveness timestamp to ensure that ad delivery is dynamically adjusted relative to the user's viewing progress, rather than being delivered at a fixed time.

[0034] The ad delivery time is calculated using the user's viewing start time and the peak timestamp for ad reception, forming a precise delivery interval. The calculation method uses the offset of the peak timestamp relative to the viewing start time as the delivery delay. For example, if the viewing start time is 10:00 and the peak timestamp is 10:17, the ad delivery interval can be set within a window of 10:16–10:18 to cover peak attention periods and maximize exposure probability. For users with multiple peak times, the first two peak times can be selected to generate the delivery interval, achieving multi-point ad coverage. The delivery interval can be further optimized by adjusting the window length (e.g., ±1 minute or ±2 minutes) to balance actual network latency and ad loading time.

[0035] In this embodiment, the specific steps of step S5 are as follows: Based on viewing behavior data streams, the real-time viewing time of users is tracked. Based on the advertising placement interval, the real-time viewing time of the user is monitored. When the user's viewing time reaches the advertising placement interval, the advertising placement mechanism is triggered, the product push sequence is embedded and the next video is pushed, and advertising push feedback information is collected. The advertising delivery range is dynamically adjusted based on the feedback information from the advertising push.

[0036] In this embodiment, the user's real-time viewing time is obtained by continuously monitoring the viewing behavior data stream. The data stream includes the user's entry time into the live stream, video playback status, pause and resume events, and timestamps for each interactive behavior. The time the user enters the live stream is recorded as the viewing start point; then, changes in the video playback status are tracked, and the effective viewing time is calculated, i.e., interrupted segments due to brief switching or accidental touches are excluded. The viewing time of each video segment is accumulated to form a real-time viewing time metric. To ensure real-time performance, a timestamp update method with a unit of 1 second can be used, incrementally calculating the user's viewing time every second, and smoothing outliers, such as zero-second intervals caused by short-term network interruptions, are not included in the total time. The advertising triggering mechanism matches the user's real-time viewing time with a preset advertising interval. When the user's current viewing time enters the advertising interval, for example, the advertising interval is 17 to 18 minutes after the start of viewing, and real-time monitoring detects that the user has watched for 17 minutes and 30 seconds, an advertisement is immediately triggered. During ad delivery, the generated product push sequence is embedded in the recommendation slot of the next video or the current video, displayed through video stream overlay, pop-ups, or interstitials. Ad push feedback is collected, including whether users viewed the ad, clicked on the product, added it to their cart, or performed other interactive actions. This feedback is recorded using timestamps and behavior types to form an ad performance dataset. Furthermore, short monitoring windows can be set, such as within ±5 seconds, to ensure ad triggering coincides with peak user attention, improving exposure and conversion potential.

[0037] The dynamic adjustment of ad delivery intervals is based on actual user feedback and behavioral responses. Collected ad push feedback is quantified, including metrics such as click-through rate (CTR), add-to-cart rate, view completion rate, and dwell time, and mapped to interval performance scores. For example, for a given ad interval, if the CTR is below 0.3 or the view completion rate is below 50%, the interval has low acceptance; conversely, if the CTR exceeds 0.6 and the dwell time is close to the interval length, the interval is considered a high-acceptance period. Adjustment strategies can employ sliding window or iterative optimization methods, advancing or delaying low-performance intervals by 1–3 minutes, while retaining or expanding high-performance intervals, ensuring that ad delivery is always synchronized with peak user attention. Ad delivery intervals can be dynamically updated based on real-time changes in user behavior; for example, if users progress faster than average, the interval is triggered earlier; if they progress slower, the interval is delayed. Through this closed-loop dynamic adjustment, ad delivery strategies can be continuously optimized, maximizing exposure and user interaction efficiency.

[0038] In this embodiment, the specific steps for embedding the product push sequence and pushing the next video are as follows: Calculate the duration of the advertising campaign period; Define the push duration for a single product; Based on the time length, the number of pushable items is calculated for the push duration to obtain the number of pushable items. Based on the number of pushable products, the product push sequence is extracted sequentially to obtain the product list to be pushed. The list of products to be pushed is processed to generate individual product advertising materials, resulting in multiple product display videos; Embed multiple product display videos and push the next video.

[0039] In this embodiment, the duration of the advertising placement interval is calculated based on the start and end times of the advertising placement determined from the user viewing behavior data stream. The start and end timestamps of the advertising placement interval are extracted; for example, the start time is 17 minutes and 10 seconds after viewing begins, and the end time is 17 minutes and 50 seconds. The interval length is calculated by subtracting these times, resulting in an available advertising duration of 40 seconds. This duration serves as a fundamental parameter for advertising placement design, determining the display duration of individual products, the number of pushes, and the playback rhythm. The interval length can be dynamically adjusted based on user viewing behavior. For example, intervals with long user dwell times and high activity levels can be extended by 1-5 seconds to increase advertising display opportunities, while intervals with interrupted viewing or decreased interaction are shortened to ensure that advertising placement is synchronized with peak user attention. The duration calculation can employ second-level precision, combined with a sliding window method to dynamically segment continuous high-activity intervals, ensuring that the advertising placement interval accurately covers the user's potential peak reception period. The push duration of a single product is set based on the length of the advertising placement interval, the complexity of the product content, and the duration of user attention. Generally, the display time for each product can be set between 5 and 15 seconds to ensure users have sufficient time to browse product information and interact, without causing viewing fatigue. For products with a large amount of product information or videos containing demonstration scenes, the display time can be appropriately extended; for products with a single image or a short text description, the display time can be shortened. The choice of push duration should also consider the total length of the advertising interval. For example, if the advertising interval is 40 seconds and the push duration for a single product is 8 seconds, then 5 products can be arranged within this interval. When setting the push duration, user viewing tolerance cycle and attention duration indicators can be combined to ensure that each product is displayed during the peak of user attention, thereby increasing exposure and click-through conversion probability.

[0040] The number of products that can be pushed is calculated by dividing the length of the ad campaign interval by the duration of each individual product's display. For example, if the ad campaign interval is 40 seconds and each product's display duration is 8 seconds, then the number of products that can be pushed is 40 ÷ 8 = 5. The calculation needs to consider buffer time and transition effects. For example, if a 1-second switching time is allowed between each product, then the number of products that can be pushed is [40 ÷ (8+1)] = 4. This method ensures that the ad campaign interval can fully accommodate product displays without exceeding the time limit or missing any. The calculation of the number of products that can be pushed can also be combined with dynamic changes in user attention. For example, an additional number of products can be displayed during high-activity intervals, while the number can be reduced during periods of fatigue, maximizing the user experience and conversion efficiency of the ad campaign. The product push sequence is generated based on suitability and potential purchase ratings, with higher-priority products listed first. Based on the number of products that can be pushed, the corresponding number is extracted sequentially from the product push sequence to form a list of products to be pushed. For example, if the number of products that can be pushed is 4, then the first 4 products are extracted from the sequence as the products to be pushed. During the sequential extraction process, a product category diversity strategy can be incorporated to avoid the consecutive appearance of similar products. The display order can be adjusted to increase the exposure opportunities for products with different attributes. The product list to be pushed includes product ID, push duration, display scenario, and advertising material information. Each record corresponds to one product, providing complete data for the next step of advertising material generation and video embedding. Corresponding advertising material videos are generated for each product in the product list to be pushed. Material generation includes integrating images or video clips, overlaying text descriptions, presenting price and promotional information, and visualizing the display scenario. The length of each product's advertising video is set according to the aforementioned push duration for a single product, typically 5–15 seconds. The material generation process is automated through a template-based approach, such as pre-designing multiple video templates and filling product information with placeholders to achieve batch generation. The generated videos include product display screens, functional feature descriptions, scene backgrounds, and visual guidance elements. Dynamic transition effects can be added to ensure smooth viewing. The generated videos can be compressed and optimized to ensure seamless embedding in the live stream without affecting playback quality.

[0041] The ad creative video is embedded into the next video in the live stream content flow for delivery. Embedding can be done via interstitial playback or overlay with recommended placements, ensuring a smooth transition between the product video and the main video content, avoiding disruption to the viewing experience. During the delivery process, the generated videos are played sequentially according to the individual product push duration and the ad interval length. The start and end times of each product video's playback, user viewing status, and interaction behavior are recorded for subsequent ad performance evaluation. The embedding strategy can be dynamically adjusted based on peak user attention and real-time viewing duration. For example, visual guidance elements and interactive buttons can be added during peak attention periods to improve click-through rates.

[0042] In this embodiment, the specific steps for dynamically adjusting the advertising delivery interval based on advertising push feedback information are as follows: Based on the feedback information from the ad push, the user's ad viewing time, product click-through rate and purchase success rate, frequency of quick swiping behavior and number of times the user stays to rewatch are calculated to obtain feedback behavior data. Calculate actual ad acceptance based on feedback behavior data; The deviation is calculated based on the actual advertising acceptance rate and the preset advertising acceptance rate to obtain the acceptance rate deviation value; The acceptance deviation value is used to evaluate the effectiveness of the campaign and generate an evaluation coefficient for the effectiveness of the campaign. The advertising placement range is dynamically adjusted based on the evaluation coefficient of the placement effect.

[0043] In this embodiment, user viewing behavior records during ad playback are converted into ad viewing duration, i.e., the cumulative dwell time from the start of the ad video to the user switching videos or closing the window, which can be counted with precision at the second or frame level, excluding short-term accidental touches. For example, viewing that is quickly abandoned by a user for less than 2 seconds is not included in the total duration. Secondly, the product click-through rate is calculated by recording the ratio of the number of product click events in the ad to the number of ad impressions, thus obtaining the click probability. The purchase success rate is calculated based on the ratio of the number of times a user completes a transaction or adds an item to the shopping cart after clicking, reflecting the ad conversion effect. The frequency of quick swiping behavior is obtained by counting the number of times users trigger fast forward or swipe to skip during ad playback, while the number of times users stay to watch the ad is counted by recognizing the number of times users go back to watch the ad after it has started playing. All these behavioral indicators are integrated into multi-dimensional feedback behavior data, forming specific characteristics of user response to ad content. The actual ad acceptance rate is generated by comprehensively evaluating the feedback behavior data. Various metrics are normalized. For example, the ratio of ad viewing time to total ad duration, product click-through rate, purchase success rate, frequency of quick swipes, and number of rewatches are converted into scores of 0–1. The frequency of quick swipes is used as a negative indicator, and the number of rewatches is used as a positive indicator. A weighted composite method is used to calculate the actual ad acceptance rate, where the weights can be set as follows: viewing time 0.4, click-through rate 0.3, purchase success rate 0.2, frequency of quick swipes -0.1, and number of rewatches 0.1. The weighted formula is as follows: Actual Acceptance Rate = 0.4 × Viewing Time Ratio + 0.3 × Click-through Rate + 0.2 × Purchase Success Rate - 0.1 × Frequency of Quick Swipes + 0.1 × Number of Rewatches. The result is a value in the range of 0–1, with the closer to 1 indicating a higher level of user acceptance of the ad content.

[0044] The acceptance deviation value is calculated by comparing the actual ad acceptance with the preset target acceptance. The preset ad acceptance is typically set based on historical ad data and user profiles. For example, a target acceptance of 0.7 means that users are expected to effectively view and interact with the ad at least 70% of the time. The deviation value is calculated as: Deviation Value = Actual Ad Acceptance – Preset Acceptance. For example, if the actual acceptance is 0.55, the deviation value is 0.55 – 0.7 = -0.15, indicating that the ad campaign is below the expected target. The deviation value can be positive or negative. A positive value indicates that the actual acceptance exceeds expectations, suggesting increasing the ad frequency or extending the ad period; a negative value indicates that the actual acceptance is below expectations, requiring optimization of ad content or adjustment of the ad time period. The campaign effectiveness evaluation coefficient is formed by quantifying the acceptance deviation value. Based on the absolute magnitude and direction of the deviation value, it is mapped to a performance rating range of 0–1. For example, a deviation within ±0.05 is defined as high performance (coefficient ≥ 0.8), a deviation within ±0.1–0.2 as medium performance (coefficient 0.5–0.7), and a deviation exceeding 0.2 as low performance (coefficient ≤ 0.3). Secondly, the performance coefficient is weighted and adjusted by combining the ad interval length, user viewing tolerance period, and historical interaction data. For example, the deviation value of highly active users increases the influence of the coefficient by 20%, while the influence of inactive users decreases by 10%, ensuring that the evaluation results reflect actual user engagement. A performance evaluation coefficient matrix is ​​generated, with each record containing user ID, ad interval, deviation value, and evaluation coefficient, which can be used for batch analysis of multiple users and multiple ads.

[0045] The ad placement intervals are dynamically adjusted through a closed-loop optimization based on the ad performance evaluation coefficient. For intervals with low evaluation coefficients (e.g., ≤0.3), the start time of the interval can be delayed or advanced, the length of the ad interval can be adjusted, or the product push sequence can be changed to increase user attention and acceptance. For example, the original 40-second interval can be shortened to 30 seconds or the interval can be moved forward by 2–3 minutes to avoid user fatigue. For intervals with high evaluation coefficients (≥0.8), the interval length can be extended, the number of product displays can be increased, or the ad can be repeated to improve the overall conversion rate. Dynamic adjustments can be combined with a sliding window strategy to track changes in user viewing time, interaction behavior, and acceptance in real time, and fine-tune the ad intervals to achieve continuous optimization. After adjustment, the ad placement plan is updated and a new ad interval list is generated, forming a cyclical optimization mechanism that ensures that ad pushes continuously match user behavior characteristics and preferences, improving conversion rates and viewing experience.

[0046] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0047] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein are implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for pushing product advertisements in live-streaming e-commerce, characterized in that, Includes the following steps: Step S1: Collect viewing behavior data stream; perform feature recognition based on viewing behavior data stream to generate viewing status description data; Step S2: Perform interest and preference analysis based on viewing status description data to obtain user preferred content data; Step S3: Calculate and match product suitability based on user preference content data to generate a product push sequence; Step S4: Analyze ad delivery time based on viewing behavior data stream and output ad delivery intervals; Step S5: Embed the product push sequence according to the advertising delivery interval.

2. The live-streaming e-commerce product advertising push method according to claim 1, characterized in that, The specific steps of step S1 are as follows: The live streaming platform monitors the user's viewing process in real time and collects viewing behavior data streams, including the distribution of dwell time, changes in the semantic rhythm of bullet comments, the density of interactive operations, and the frequency of screen switching. Perform time-series serialization processing on the viewing behavior data stream to construct the viewing behavior trajectory; Based on the viewing behavior trajectory, the viewing stability, interaction response rate, and attention duration are calculated to generate viewing state description data.

3. The live-streaming e-commerce product advertising push method according to claim 1, characterized in that, The specific steps of step S2 are as follows: Identify multiple live streaming rooms based on viewing behavior data streams; The live stream content of the multiple live streaming rooms is analyzed to obtain multiple content tags; Calculate user activity levels in different live streaming rooms based on viewing behavior data streams; Based on the behavioral activity level, multiple content tags are associated and labeled to obtain the interaction intensity coefficient of different live streaming rooms; Based on the viewing status description data and the interaction intensity coefficient, interest preference analysis is performed to obtain user preferred content data.

4. The live-streaming e-commerce product advertising push method according to claim 1, characterized in that, Step S3 is as follows: Extract the product database; parse the product attributes of each product in the product database and output the attribute information of all products; The attribute information includes product category, price range, functional characteristics, and display scenarios; Based on user preference content data, the attribute information is matched and multiple suitable products are extracted. Predict purchase intentions for multiple compatible products and generate potential purchase scores for different products; The product recommendation sequence is generated by performing gradient sorting based on potential purchase scores.

5. The live-streaming e-commerce product advertising push method according to claim 1, characterized in that, The specific steps of step S4 are as follows: Obtain user identity information; perform identity clustering analysis based on user identity information to obtain user type characteristics; Based on the viewing behavior data stream, the viewing tolerance cycle, fatigue time point and peak activity period are calculated to obtain the user's viewing rhythm information; Based on user viewing rhythm information and user type characteristics, conduct in-depth user evaluation and construct user feature profiles; Based on user profile characteristics, predict ad placement and output ad placement ranges.

6. The live-streaming e-commerce product advertising push method according to claim 5, characterized in that, The specific steps for predicting ad placement based on user feature profiles and outputting the optimal ad placement range are as follows: Based on user feature profiles, predict the probability of ad acceptance and generate an ad acceptance probability curve; Peak points of acceptance were marked based on the advertising acceptance probability curve; Calculate the timestamp of the peak acceptance point; Determine the user's viewing start timestamp based on the viewing behavior data stream; The advertising time is calculated based on the user's viewing start timestamp and the timestamp, and the advertising placement interval is output.

7. The live-streaming e-commerce product advertising push method according to claim 1, characterized in that, The specific steps of step S5 are as follows: Based on viewing behavior data streams, the real-time viewing time of users is tracked. Based on the advertising placement interval, the real-time viewing time of the user is monitored. When the user's viewing time reaches the advertising placement interval, the advertising placement mechanism is triggered, the product push sequence is embedded and the next video is pushed, and advertising push feedback information is collected. The advertising delivery range is dynamically adjusted based on the feedback information from the advertising push.

8. The live-streaming e-commerce product advertising push method according to claim 7, characterized in that, The specific steps for embedding the product push sequence and pushing the next video are as follows: Calculate the duration of the advertising campaign period; Define the push duration for a single product; Based on the time length, the number of pushable items is calculated for the push duration to obtain the number of pushable items. Based on the number of pushable products, the product push sequence is extracted sequentially to obtain the product list to be pushed. The list of products to be pushed is processed to generate individual product advertising materials, resulting in multiple product display videos; Embed multiple product display videos and push the next video.

9. The live-streaming e-commerce product advertising push method according to claim 7, characterized in that, The specific steps for dynamically adjusting the advertising delivery range based on advertising push feedback information are as follows: Based on the feedback information from the ad push, the user's ad viewing time, product click-through rate and purchase success rate, frequency of quick swiping behavior and number of times the user stays to rewatch are calculated to obtain feedback behavior data. Calculate actual ad acceptance based on feedback behavior data; The deviation is calculated based on the actual advertising acceptance rate and the preset advertising acceptance rate to obtain the acceptance rate deviation value; The acceptance deviation value is used to evaluate the effectiveness of the campaign and generate an evaluation coefficient for the effectiveness of the campaign. The advertising placement range is dynamically adjusted based on the evaluation coefficient of the placement effect.

Images