A live e-commerce user behavior analysis and accurate recommendation method

Abstract

The application discloses a live broadcast e-commerce user behavior analysis and accurate recommendation method, and relates to the technical field of live broadcast e-commerce intelligent recommendation.The specific steps of the method are as follows: collecting user historical consumption static data and live broadcast room dynamic behavior data and fusing the same to output a comprehensive interest quantization value; analyzing a commodity life cycle and quantizing user three-dimensional interest to form a comprehensive parameter set; formulating a personalized recommendation scheme, determining a display form and content; executing recommendation and collecting interactive data; processing the interactive data to generate supplementary data, returning and updating to realize a closed loop iteration; the application can mine user interest in multiple dimensions, comprehensively and accurately evaluate, avoid recommendation deviation, fit user demand, improve interactive willingness and shopping experience, help merchants accurately connect with customers and improve conversion efficiency; the application can design differentiated recommendation forms, customize targeted content, build a closed loop iteration mechanism, dynamically optimize recommendation strategies, and enhance user stickiness and platform competitiveness.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation technology for live-streaming e-commerce, specifically a method for analyzing user behavior and making accurate recommendations in live-streaming e-commerce. Background Technology

[0002] With the rapid development of the digital economy, live-streaming e-commerce, with its advantages of real-time interaction and immersive scenarios, has become a significant growth driver in the e-commerce industry, with user and transaction volumes continuing to rise. Accurate recommendation, as a key element in improving user experience and optimizing merchant conversion, relies on in-depth analysis of user behavior data to uncover genuine needs. User behavior in live-streaming e-commerce scenarios encompasses both static data such as historical consumption and dynamic data such as clicks, dwell time, and interactions within the live stream. These two types of data together constitute a complete profile of user interests. How to efficiently integrate and utilize this data to achieve accurate recommendations has become a crucial issue that the industry urgently needs to address.

[0003] Existing live-streaming e-commerce recommendation technologies suffer from several shortcomings: Data fusion methods are often crude, relying heavily on static historical consumption data and neglecting the value of real-time dynamic behavior in the live stream. Even when attempting to integrate dynamic data, these methods fail to consider the time decay characteristics of behavior, resulting in unscientific weight allocation and significant biases in the overall assessment of user interests. Interest quantification dimensions are limited, focusing primarily on category preferences and failing to delve into users' core needs regarding product functionality and usage scenarios. Furthermore, recommendation strategies are not adjusted based on product lifecycle stages, leading to insufficient adaptability. Recommendation display formats are rigid, lacking differentiated interface and content design based on recommendation suitability, resulting in low user reception efficiency. Finally, a complete closed-loop feedback mechanism is lacking, preventing real-time updates to the model from user interaction data, causing recommendation iteration delays and making it difficult to adapt to dynamically changing user needs.

[0004] In summary, existing recommendation methods suffer from low accuracy, poor user experience, and insufficient merchant conversion rates due to issues such as unscientific data fusion, incomplete interest quantification dimensions, rigid adaptation strategies, and a lack of feedback loops. These shortcomings make it difficult to meet the personalized and real-time recommendation needs of the live-streaming e-commerce industry. To address these challenges, there is an urgent need to develop a precise recommendation method that can systematically integrate static and dynamic data, quantify user interests across multiple dimensions, optimize adaptation strategies based on the product lifecycle, and continuously improve through closed-loop iteration. This would enhance the targeting and effectiveness of recommendations and promote the high-quality development of the live-streaming e-commerce industry. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for analyzing and accurately recommending user behavior in live-streaming e-commerce. This method collects historical static consumption data and dynamic behavior data from the live-streaming room, and then uses time-series weighted fusion to obtain a comprehensive interest quantification value; it analyzes three dimensions of interest, including product lifecycle and quantified user category, to form a comprehensive parameter set; it uses a multi-dimensional adaptation algorithm to formulate a personalized recommendation scheme; it displays and collects interactive data through a real-time recommendation engine; and it classifies, quantifies, and updates the data to achieve closed-loop iteration.

[0006] To solve the above-mentioned technical problems, this invention provides the following technical solution: a method for analyzing and accurately recommending user behavior in live-streaming e-commerce, the specific steps of which are as follows:

[0007] S100. Data Acquisition and Fusion: Collect historical user consumption data and perform static feature quantification to obtain static feature quantified data; simultaneously collect dynamic behavior data of users in the live broadcast room and analyze it to obtain dynamic behavior analysis data; use a time-series weighted fusion algorithm to perform weighted fusion of the static feature quantified data and the dynamic behavior analysis data to output the user's comprehensive interest quantification value.

[0008] S200, Comprehensive Parameter Generation: Based on the quantitative value of user comprehensive interest, analyze the online duration, exposure, sales trend and inventory status of each product in the live broadcast room, determine the product life cycle stage and determine the adaptation coefficient; use three-dimensional layered quantification technology to layer and quantify user interest in three dimensions: category, function and scenario, and generate three-dimensional quantitative values ​​of user interest; integrate the adaptation coefficient and the three-dimensional quantitative values ​​of user interest to form a comprehensive parameter set;

[0009] S300, Recommendation scheme formulation: Receive a comprehensive parameter set, calculate the comprehensive fit degree using a multi-dimensional adaptation algorithm; recommend product display formats based on the comprehensive fit degree, customize scenario-based recommendation content, and integrate them to form a personalized recommendation scheme;

[0010] S400, Recommendation Execution: Receive personalized recommendation schemes, and use the real-time recommendation engine in the live broadcast room to display recommended content to target users in predetermined interface positions in the product display area, bullet screen interaction bar, or side floating window, and collect interactive behavior data of target users in response to the recommended content in real time;

[0011] S500, Closed-loop iteration: Receive the interactive behavior data of the target user, classify and quantify the interactive behavior data according to the preset classification rules, and generate dynamic behavior supplementary data; send the dynamic behavior supplementary data back to the data acquisition and fusion step to update the static feature quantification data and dynamic behavior analysis data.

[0012] Furthermore, the process of collecting users' historical consumption data and performing static feature quantification to obtain static feature quantified data is as follows: Collect users' historical consumption data on the live-streaming e-commerce platform; clean the collected historical consumption data to obtain standardized static data; perform static feature quantification on the standardized static data to generate static feature quantified data; the historical consumption data includes historical purchased product categories, purchase frequency, average order value, product favorites, historical refund records, and repurchase frequency. The process of simultaneously collecting and parsing users' dynamic behavior data within the live-streaming room to obtain dynamic behavior parsing data is as follows: Simultaneously collect users' dynamic behavior data within the live-streaming room; classify the collected dynamic behavior data, distinguishing between structured and unstructured behavior data; quantify the structured behavior data and parse the unstructured content; integrate the parsed structured and unstructured behavior data to generate dynamic behavior parsing data; the dynamic behavior data includes live-streaming room dwell time, product click count, bullet screen question content, product detail page dwell time, live-streaming room likes and shares, and temporary add-to-cart actions.

[0013] Furthermore, the calculation formula for the time-weighted fusion algorithm is as follows: ,in, Quantify the user's overall interests; The static feature weight ratio coefficient; Quantify data for static features; This represents the total number of user dynamic behaviors collected within the live stream. As the basis for dynamic behavior weights; This is the time decay coefficient; For the first The time difference between the occurrence time of a dynamic behavior and the current calculation time; For the first The system analyzes dynamic behavior data and, based on time-series weighted calculation logic, integrates the static features corresponding to users' historical consumption with the real-time dynamic behavior features in the live broadcast room. By allocating reasonable weights to the two types of features, it achieves an organic combination and outputs a comprehensive quantitative value that reflects the user's current real shopping interest.

[0014] Furthermore, the rules for determining the product lifecycle stage and matching coefficient in the live stream are as follows: Analyze the online duration, exposure, sales trends, and inventory status data of each product in the live stream, divide the product lifecycle into a preheating period, a growth period, a peak period, and a decline period, match it to the corresponding lifecycle stage, and assign a unique matching coefficient to each stage; the preheating period is defined as online duration ≤ 2h, exposure growth rate < 10%, sales growth rate < 5%, and inventory remaining rate ≥ 90%; the growth period is defined as 2h < online duration ≤ 8h, 10% ≤ exposure growth rate < 50%, 5% ≤ sales growth rate < 80%, and 70% ≤ inventory remaining rate < 90%; the peak period is defined as 8h < online duration ≤ 24h, exposure growth rate within ±10%, sales growth rate within ±5%, and 30% ≤ inventory remaining rate < 70%; the decline period is defined as online duration > 24h, exposure growth rate < -10%, sales growth rate < -20%, and inventory remaining rate < 30%.

[0015] Furthermore, the specific steps for using three-dimensional layered quantification technology to layer and quantify user interests across three dimensions—category, function, and scenario—to generate three-dimensional quantified values ​​of user interests are as follows: Combining the user's comprehensive interest quantification value with historical user consumption data and livestream dynamic behavior data, determine the user's category interest direction; based on the determined category interest direction, extract the functional attributes of products clicked, added to cart, favorited, and purchased by the user within that category to extract the user's functional interest preferences; match the real-time scenario of the livestream, combining the user's past consumption behavior in the corresponding scenario with behavioral data within the livestream scenario to determine the user's scenario interest tendency; quantify the three dimensions of category interest direction, functional interest preference, and scenario interest tendency to generate a three-dimensional quantified value of user interests that includes the quantification results of the category, function, and scenario dimensions.

[0016] Furthermore, the calculation formula for the multi-dimensional adaptation algorithm is as follows: ,in, For overall compatibility; Product compatibility coefficient; This represents the weighting coefficient for the category interest dimension. Quantify user category interests; For the functional interest dimension weight coefficient; Quantify user interest in features; For the scene interest dimension weight coefficient; Quantify user interests in specific scenarios; It quantifies the overall user interest value; by organically integrating multi-dimensional features, it quantifies the suitability between products and users, providing accurate quantitative judgment basis for subsequent priority delineation of recommendation time windows, selection of recommendation display formats, and customization of scenario-based recommendation content, thereby improving the accuracy of matching recommendation solutions with user needs.

[0017] Furthermore, the process of integrating to form a personalized recommendation scheme is as follows: Three intervals are divided based on overall fit: a high fit interval (overall fit ≥ 0.7), a medium fit interval (0.4 ≤ overall fit < 0.7), and a low fit interval (overall fit < 0.4). The recommended product display area in the high fit interval is a full-screen pop-up card format; the recommended interactive bar in the medium fit interval is a text and image link format; and the recommended side floating window in the low fit interval is a simple card format. Combining the three-dimensional quantitative results of user interests associated with the corresponding intervals, product attributes, and the real-time scene of the live stream, customized scenario-based recommendation content that highlights interest and scenario fit is created. The recommendation display format, scenario-based recommendation content, display duration, and display position are integrated to form a personalized recommendation scheme.

[0018] Furthermore, the real-time recommendation engine includes a data receiving module, a resource scheduling module, a display control module, and a behavior collection module, with the following specific structure: The data receiving module receives personalized recommendation schemes and parses the display format, content, location, and target user information; the resource scheduling module coordinates with each display interface in the live broadcast room to allocate display resources; the display control module controls the loading, display duration, and timing of content at predetermined locations according to the recommendation scheme; and the behavior collection module captures real-time interactive behavior data of target users, such as clicking, pausing, and closing the recommended content.

[0019] Furthermore, the preset classification rules are based on interactive behavior data to construct dimensions; the interactive behavior data is classified and quantified according to the preset classification rules to generate dynamic behavior supplementary data. The specific process is as follows: the collected interactive behavior data is classified into click type, dwell type, and operation type and the classification is completed; each type of behavior is quantified and converted into statistically verifiable quantitative processing results, and the quantitative processing results of the three types of behavior are integrated into dynamic behavior supplementary data.

[0020] Compared with existing technologies, this method for analyzing and accurately recommending user behavior in live-streaming e-commerce has the following advantages:

[0021] I. This invention integrates users' historical static consumption data with real-time dynamic behavior data from live streams, employing a scientific fusion method to process the correlation between the two types of data. Simultaneously, it delves into users' core interests across three dimensions: category, function, and scenario. By adjusting adaptation strategies based on the characteristics of products at different lifecycle stages, it achieves a comprehensive and accurate assessment of user interests. This multi-dimensional and in-depth demand mining approach effectively avoids recommendation biases caused by single data sources or one-sided interest dimensions, making recommended content more aligned with users' actual consumption needs. This enhances users' willingness to interact and their shopping experience in live streams, helps merchants accurately connect with target customer groups, reduces ineffective exposure, improves product conversion efficiency, and builds a more efficient supply and demand matching ecosystem for the platform.

[0022] Second, this invention designs differentiated recommendation display formats based on comprehensive adaptability, and customizes targeted recommendation content according to the real-time scenario of the live broadcast room, ensuring that the recommendation information is presented in a way that is easily accepted by users. At the same time, it establishes a complete closed-loop iteration mechanism, collects user interaction data on recommended content in real time, performs classification and quantification processing, and sends it back to the data collection and integration stage for data updates. This dynamic optimization mode allows the recommendation strategy to adapt to changes in user interests, breaking the limitations of traditional recommendation methods that are rigid and have lagging iterations, improving the timeliness and accuracy of recommended content, helping the platform to continuously optimize the recommendation algorithm, enhancing user stickiness and platform competitiveness, and providing strong support for the sustainable development of the live broadcast e-commerce industry.

[0023] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0025] Figure 1 A flowchart illustrating the steps of a live-streaming e-commerce user behavior analysis and precise recommendation method;

[0026] Figure 2 A flowchart for a method of analyzing and accurately recommending user behavior in live-streaming e-commerce;

[0027] Figure 3 This is a schematic diagram illustrating the data transmission of comprehensive parameters of livestream users in a livestream e-commerce user behavior analysis and precise recommendation method. Detailed Implementation

[0028] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0029] Example 1: Collect historical consumption data of users' past purchases of beauty products on live-streaming e-commerce platforms. Clean the collected historical consumption data, removing duplicate records, errors such as abnormally priced purchases, and invalid information such as purchase data from accounts not belonging to the user. This makes the standardized static data more reliable, ensuring that the generated static feature quantification data truly reflects users' past consumption preferences. Simultaneously, collect dynamic behavior data of users within the live-streaming room. Classify and process this dynamic behavior data, clearly distinguishing between structured behavior data such as clicks, adding to cart, and favorites, and unstructured behavior data such as bullet screen content and interactive comments. Structured behavior data has clear operational directions, facilitating direct quantitative analysis, while unstructured behavior data hides deeper user needs and questions, requiring analysis to extract valuable information. Structured behavioral data is quantified, converting different types of operations into statistically measurable values, allowing for precise measurement of the intensity and frequency of behavior. Unstructured bullet screen content is parsed, extracting keywords such as "makeup staying power," "moisturizing," "sensitive skin," and "whitening effect" to identify users' core needs and potential concerns. The parsed structured and unstructured behavioral data are then integrated to generate dynamic behavioral analysis data, ensuring the data contains both explicit operational information and deeper expression of needs, comprehensively presenting users' behavioral characteristics within the live stream. Subsequently, a time-series weighted fusion algorithm is used to weight and fuse the static feature quantification data and the dynamic behavioral analysis data. The formula is: ,in, Quantify the user's overall interests; The static feature weight ratio coefficient; Quantify data for static features; This represents the total number of user dynamic behaviors collected within the live stream. As the basis for dynamic behavior weights; This is the time decay coefficient; For the first The time difference between the occurrence time of a dynamic behavior and the current calculation time; For the first This algorithm analyzes dynamic behavior data and fully considers both the stable preferences formed by users' historical consumption and the instantaneous interests generated in the live broadcast room. By reasonably balancing the weights of the two and combining the time decay characteristics of dynamic behavior, it ensures that the quantitative value of users' comprehensive interests does not deviate from the consumption habits formed by users over a long period of time, but also accurately matches the instantaneous needs of the current live broadcast room scenario.

[0030] Based on the core quantitative value of comprehensive user interests, which fully integrates historical and real-time behavioral information, it can provide precise guidance for product lifecycle analysis and user interest mining. Analyzing the launch duration of various beauty products in the live stream, such as a new matte lipstick launched for 3 days and a classic cushion foundation launched for 2 months, the system tracks exposure (the number of times users see the product and its growth trend), sales trends (continuously rising, stable, or gradually declining), and inventory status (sufficient stock, tight stock, or nearing sell-out). Through comprehensive analysis of these multi-dimensional data, the product lifecycle is precisely divided into pre-launch, growth, peak, and decline phases. For example, a new matte lipstick, after 3 days, shows gradually increasing exposure but low sales, indicating it's in the market awareness stage and is considered in the pre-launch phase. At this time, the adaptation coefficient should focus on guiding users to understand the product. A classic cushion foundation, after 2 months, shows stable sales and sufficient stock, indicating high market acceptance and is considered in the peak phase; the adaptation coefficient should focus on promoting conversion. A unique adaptation coefficient is assigned to each lifecycle stage, ensuring the coefficient accurately matches the product's current promotional goals and market status, making subsequent recommendations more aligned with the product's promotional rhythm. By employing a three-dimensional stratified quantification technique, user interests are stratified and quantified across three dimensions: category, function, and scenario. Combined with comprehensive user interest quantification values ​​and linked to historical user consumption data and livestream dynamic behavior data, this approach precisely identifies user category interests concentrated on lipsticks and foundations. This process avoids generalized judgments of user interests, focusing on core consumer categories. Based on the identified category interest direction, the functional attributes of products clicked, added to cart, favorited, and purchased by users within that category are extracted in depth. For example, the lasting power, moisturizing properties, and color payoff of lipsticks, and the coverage, adherence, and oil control of foundations. This clarifies user functional interest preferences as long-lasting and moisturizing, making functional interest capture more targeted and avoiding recommendations of products that do not match user functional needs. Matching the live stream to real-time daily makeup tutorials, where users are more concerned with the product's application effect in actual makeup, and combining this with users' past consumption behavior in daily commuting scenarios (such as a preference for convenient and long-lasting beauty products), and behavioral data in the live stream scenario (such as users frequently asking about product performance during commutes), the system determines that the user's scenario interest is for daily commuting makeup combinations. This ensures that the determination of scenario interest is closely integrated with the current live stream scenario and the user's past consumption habits, making interest capture more aligned with actual usage needs. The system quantifies three dimensions of user interest: category interest, functional interest preference, and scenario interest tendency, generating a three-dimensional quantified value of user interest that comprehensively and accurately presents the intensity and direction of user interest across different dimensions. The system integrates the matching coefficient with the three-dimensional quantified value of user interest to form a comprehensive parameter set.

[0031] The system receives a comprehensive parameter set, which includes product fit coefficients and three-dimensional user interest quantification values, providing a comprehensive and accurate data foundation for calculating the overall fit score. The overall fit score is then calculated using a multi-dimensional fit algorithm, with the following formula: ,in, For overall compatibility; Product compatibility coefficient; This represents the weighting coefficient for the category interest dimension. Quantify user category interests; For the functional interest dimension weight coefficient; Quantify user interest in features; For the scene interest dimension weight coefficient; Quantify user scenario interests; The algorithm quantifies users' overall interests. It systematically integrates product fit coefficients with user interest quantification values ​​across three dimensions: category, function, and scenario. Through scientific weighting and calculation logic, it accurately measures the degree of matching between products and user needs, avoiding recommendation bias caused by single-dimensional judgments. This ensures that the overall fit score truly reflects the product's attractiveness to users. The algorithm categorizes products into high, medium, and low fit scores, each with a different recommendation display format to ensure the format matches the degree of matching. Products in the high fit score have the highest degree of alignment with user needs, and are displayed as full-screen pop-up cards in the product display area. This visually impactful format quickly attracts user attention, allowing them to immediately focus on highly relevant products. The medium fit score recommends text and image links in the interactive comment section, providing product information without interfering with the user's live stream viewing. The low fit score recommends simple cards in a side-mounted floating window to avoid excessive user interference. When a user's overall fit is in the high-fit range, based on the three-dimensional quantitative results of user interests associated with this range, the attributes of beauty products, and the real-time scenario of daily makeup tutorials in the live stream, customized scenario-based recommendation content is created that highlights the long-lasting moisturizing function and suits daily commuting scenarios. This includes product wear time test results, such as wear performance under different environmental temperatures and the makeup condition after an 8-hour commute; explanations of moisturizing ingredients, such as the mechanisms of action of added hyaluronic acid and ceramides; and makeup examples that match everyday outfits, such as different makeup effects with professional and casual wear. This content allows users to intuitively feel the high degree of fit between the product and their own needs, quickly understand the core advantages and usage value of the product, and reduce the user's decision-making cost. The recommendation display format, scenario-based recommendation content, display duration, and display position are integrated to form a personalized recommendation solution.

[0032] The system receives personalized recommendation plans, which clearly define key elements such as display format, content, location, and duration, providing clear guidance for the operation of the real-time recommendation engine. Recommendations are then made through the real-time recommendation engine within the live stream. This engine consists of a data receiving module, a resource scheduling module, a display control module, and a behavior collection module. Each module has a clear division of labor and works collaboratively to ensure efficient and accurate recommendation execution. The data receiving module receives the personalized recommendation plan and accurately analyzes its display format (full-screen pop-up cards), content (beauty products highlighting long-lasting hydration and daily commuting suitability), location (product display area), and target user information (users with high overall suitability). The analysis process strictly adheres to the plan's requirements to avoid information bias leading to recommendation errors. The resource scheduling module coordinates with the live stream's product display area to allocate display resources. Based on the current traffic and load status of the live stream, it rationally allocates server and bandwidth resources to ensure smooth loading of the full-screen pop-up card recommendation content without lag or delays. Simultaneously, it avoids resource competition with the live stream screen and other interactive content, ensuring a superior overall viewing experience. The display control module controls the loading, display time, and duration of recommended content in the product display area in the form of full-screen pop-up cards according to the recommendation scheme. The loading process is fast and does not affect the normal playback of the live broadcast. When displayed, the card content is clear, the colors are coordinated, and it meets the visual presentation requirements of beauty products. The duration is strictly executed according to the scheme settings, ensuring that users have enough time to browse product information without the display time affecting their viewing of the live broadcast. The behavior collection module captures the target user's interactive behavior data in real time, such as clicking, dwelling, and closing the recommended content. Click behavior includes clicking the card to view product details, clicking the purchase link, clicking the favorite button, etc. Dwell behavior includes the dwell time on the pop-up card display page, the browsing time on the product details page, etc. Closing behavior includes actively closing the pop-up card, and not performing any operation until the card closes automatically. This module can accurately and comprehensively capture every user interaction action, ensuring that user feedback data is not missed or biased, and providing real and effective data support for subsequent closed-loop iterations.

[0033] The system receives interactive behavior data from target users and categorizes and quantifies this data according to preset classification rules. These rules are built upon the interactive behavior data to construct dimensions. The collected interactive behavior data is first divided into click, dwell, and action categories and then categorized. Click categories include actions with clear targets, such as clicking pop-up cards to view details, clicking product purchase links, and clicking favorites buttons. Dwell categories include behaviors reflecting user attention levels, such as the duration of dwell time on pop-up display pages and the browsing time in different sections of the product details page. Action categories include actions with practical significance, such as adding products to the cart via pop-up cards, completing payments, and canceling favorites. The classification process strictly follows the preset rules to ensure that each type of behavior has a clear classification and that there is no overlap or omission. Each type of behavior is then quantified. Click-related behaviors are assigned corresponding quantitative values ​​based on different stages of the click process; for example, clicking on a details page has a higher quantitative value than simply clicking a card. Dwell-related behaviors are quantified based on the duration of the dwell time; longer dwell times result in higher quantitative values. Operation-related behaviors are assigned different quantitative values ​​based on the completion level and importance of the operation; for example, completing a payment has a higher quantitative value than simply adding to the cart. This quantification process ensures that the intensity of user interaction and the degree of preference can be accurately quantified and transformed into statistically comparable results. The quantification results of the three types of behaviors are integrated into dynamic behavior supplementary data. This data comprehensively summarizes user feedback on recommended content, clearly presenting the intensity of user interest, focus, and behavioral tendencies towards recommended products. The dynamic behavior supplementary data is then fed back to the data collection and fusion step to update the static feature quantification data and dynamic behavior analysis data in a timely manner. Figure 1 As shown, static feature quantification data can be incorporated into the latest changes in users' consumption preferences, while dynamic behavior analysis data can reflect the latest real-time needs of users.

[0034] Example 2: Collect historical consumption data of users' past purchases of home furnishings on live-streaming e-commerce platforms. Clean this data to remove redundant information such as duplicate purchase records and invalid orders resulting from erroneous operations. Correct errors such as incorrect size or material information. This makes the standardized static data more accurate and standardized, eliminating data impurities that interfere with subsequent processing. The resulting static feature quantification data accurately reflects users' past home furnishing consumption habits and stable preferences. Simultaneously, collect dynamic behavior data of users within the live-streaming room. Classify this dynamic behavior data, clearly distinguishing between structured behavior data with explicit operational attributes such as clicks, favorites, and adding to cart, and unstructured behavior data without a fixed format, such as bullet screen content and interactive comments. Structured behavior data directly reflects users' behavioral intentions, facilitating rapid quantitative analysis. Unstructured behavior data contains more detailed and specific expressions of user needs, requiring professional parsing and processing to extract effective information and avoid overlooking users' potential demands. Structured behavioral data is quantified, transforming different types and intensities of behavior into calculable values. For example, click behavior is assigned a quantifiable value based on the number of clicks and the click stage, while add-to-cart and favorite behaviors are given higher quantifiable weights, allowing the importance of behavior to be reflected through quantifiable results. Unstructured bullet screen content is parsed to extract core keywords such as small apartment, easy installation, large capacity, and environmental protection, identifying users' core needs and concerns. The parsed structured and unstructured behavioral data are integrated to generate dynamic behavioral analysis data. Then, a time-series weighted fusion algorithm is used to weight and fuse the static feature quantified data and the dynamic behavioral analysis data. This algorithm can reasonably balance the weight of users' historical stable preferences and current real-time interests, while considering the time decay characteristics of dynamic behavior. Recent dynamic behaviors have a greater impact on users' current interests. Through this scientific fusion method, the output user comprehensive interest quantification value can both continue the long-term consumption preferences of users and accurately match the immediate needs of the current live broadcast scenario, avoiding recommendation lag caused by relying solely on historical data or one-sided recommendations caused by relying solely on real-time data, making the user comprehensive interest quantification value more valuable.

[0035] Based on a comprehensive user interest quantification value, which fully integrates historical and real-time behavioral information, this data provides precise guidance for product lifecycle analysis and user interest mining, ensuring that subsequent parameter generation is more aligned with user needs and the actual situation of the products. Analysis includes the online duration of various home furnishing products in the live stream (e.g., a new multi-functional storage cabinet launched one week ago, a classic fabric sofa launched six months ago), exposure (the number of times a product is seen by users in the live stream and its growth trend), sales trends (e.g., continuous rise, stable, gradual decline), and inventory status (e.g., sufficient inventory, dwindling inventory, nearing sell-out). Through comprehensive analysis and judgment of these multi-dimensional data, the product lifecycle is accurately divided into the pre-heating period, growth period, peak period, and decline period. For example, a new multi-functional storage cabinet, launched for one week, saw a continuous increase in exposure and sales, with market attention gradually rising, indicating it's in the upward phase of market promotion. At this stage, the matching coefficient should focus on highlighting the product's core selling points to attract more users. Conversely, a classic fabric sofa, launched for six months, experienced stable sales but sluggish growth, with inventory gradually decreasing, indicating it's at the end of its product lifecycle and in the decline phase. In this case, the matching coefficient should focus on promoting inventory clearance and improving conversion efficiency. Assigning a unique matching coefficient to each lifecycle stage ensures the coefficient accurately matches the product's current promotional goals and market conditions, making subsequent recommendation strategies more aligned with the product's promotional pace and market demand, avoiding recommendations that are incompatible with the product's current stage. Utilizing a three-dimensional layered quantification technology, user interests are layered and quantified across three dimensions: category, function, and scenario. This technology comprehensively and deeply mines user interests from different dimensions, avoiding incomplete interest capture caused by single-dimensional analysis. Combining the user's comprehensive interest quantification value with historical user consumption data and livestream dynamic behavior data, the system accurately determines that user category interests are concentrated in storage furniture and dining tables. Based on the user's interest in this product category, we deeply extract the functional attributes of products clicked, added to cart, favorited, and purchased by users within this category. For example, storage furniture might have features like storage capacity, partition design, and ease of installation; dining tables might have features like material safety, load-bearing capacity, and ease of cleaning. This reveals that users' functional preferences are large capacity and ease of installation. Matching the live stream with a real-time small-apartment decorating scenario, where users are more concerned with product space utilization and size adaptability, and combining this with past consumer behavior in small-apartment decorating scenarios (such as a preference for space-saving and multifunctional furniture), and behavioral data from the live stream (such as frequent inquiries about product size and suitability for small apartments), we determine that the user's scenario interest is focused on small-apartment home furnishing. This ensures that the determination of scenario interests closely aligns with the current live stream scenario and the user's past consumption habits, allowing interest capture to better match the user's actual usage scenario and ensuring that recommended products truly meet the user's needs in specific scenarios.The system quantifies user interests across three dimensions: category interest, functional interest preference, and scenario interest tendency. This transforms abstract interest preferences into calculable and comparable numerical values, generating a three-dimensional quantified value of user interest that includes the quantified results for category, function, and scenario dimensions. This quantified value comprehensively and accurately presents the intensity and direction of user interests across different dimensions, making user interests intuitive and measurable. The matching coefficient is then integrated with the three-dimensional quantified value of user interests to form a comprehensive parameter set, such as... Figure 3 As shown.

[0036] The system receives a comprehensive parameter set and calculates the overall fit using a multi-dimensional adaptation algorithm. This algorithm systematically integrates the product fit coefficient with the user's three-dimensional interest quantification values ​​(category, function, and scenario) according to a scientific weight allocation logic. Based on the overall fit score, the system divides the system into high, medium, and low fit ranges, each corresponding to a different recommendation display format. This division ensures a high degree of alignment between the recommendation format and the matching degree, improving the effectiveness of the recommendations. When a user's overall fit score is in the medium fit range, the recommended products are displayed as image and text links in the interactive comment section. This display format does not affect the user's overall live streaming experience and naturally presents product information while the user browses the comments, achieving a balance between passive recommendation and active engagement, allowing users to learn about suitable products without being disturbed. By combining the three-dimensional quantitative results of user interests associated with this interval, the attributes of home furnishing products, and the real-time scenario of small apartment decoration in the live broadcast room, customized scenario-based recommendation content is created that highlights the characteristics of large-capacity storage and easy installation, making it suitable for small apartment scenarios. This includes internal storage partition diagrams of the products, clearly showing the storage functions and capacities of different areas, allowing users to intuitively understand their storage capacity; simplified installation instructions, presenting the core installation steps with concise pictures and text, reducing users' concerns about installation difficulty; and small apartment space placement effect diagrams, showing how the products can be placed and the area they occupy in different small apartment spaces, allowing users to quickly determine whether the products are suitable for their own apartment layout. This scenario-based content accurately targets users' core needs, allowing them to quickly perceive the fit between the products and their own needs, shortening the user's decision-making path. The recommendation display format, scenario-based recommendation content, display duration, and display location are integrated to form a personalized recommendation scheme.

[0037] The system receives personalized recommendation schemes and executes recommendations through a real-time recommendation engine within the live stream. This engine comprises a data receiving module, a resource scheduling module, a display control module, and a behavior collection module. Each module performs its specific function and works collaboratively to form a complete recommendation execution chain, ensuring the accuracy and efficiency of the recommendations. The data receiving module receives and precisely analyzes the personalized recommendation schemes, specifying that the display format is a text and image link in the interactive bullet screen. The recommended content highlights home furnishing products with large storage capacity, easy installation, and suitability for small apartments. The display location is the interactive bullet screen, and the target users are those with a moderate overall suitability. The analysis process strictly adheres to the scheme requirements, ensuring that every piece of information is accurately interpreted and avoiding recommendation errors due to analysis deviations. The resource scheduling module coordinates with the interactive bullet screen to allocate display resources. Based on the current bullet screen traffic and server load, it rationally allocates network bandwidth and storage resources to ensure that the text and image link recommendations load quickly and display smoothly without lag or loading failures. Simultaneously, it avoids resource competition with other bullet screen content and live stream footage, ensuring the overall stability of the live stream and the user viewing experience. The display control module controls the loading, display, and duration of recommended content in the form of image-text links in the bullet screen interaction bar according to the recommendation scheme. The loading process is seamlessly integrated into the bullet screen stream without affecting the normal scrolling of other bullet screens. During display, it ensures that the images in the image-text links are clear, the text is easy to read, and the colors are coordinated with the overall style of the bullet screen bar. The duration is strictly executed according to the scheme settings, ensuring that users have enough time to discover and click on the links to view details, while preventing links from occupying bullet screen space for too long and affecting the display of other bullet screens. The behavior collection module captures the target user's interactive behavior data in real time, such as clicking, dwelling, and closing the recommended content. Click behavior includes clicking on image-text links to enter the product details page, clicking the favorite or add to cart buttons in the link, etc. Dwell behavior includes the total browsing time on the product details page, the dwell time in different feature introduction sections, etc. Closing behavior includes actively closing the details page, not taking any action until the link expires, etc. This module can accurately and comprehensively capture every user interaction action, ensuring that user feedback information is not missed or distorted, and providing real and effective data support for subsequent closed-loop iterations.

[0038] The system receives interactive behavior data from target users and categorizes and quantifies this data according to preset classification rules. First, the collected interactive behavior data is divided into click-based, dwell-based, and action-based categories. Click-based behaviors include clicking on image / text links to enter product details pages, clicking on different functional sections within the details page, and clicking on "favorite" or "add to cart" buttons—actions with clear directional intent that directly reflect the user's initial interest in the product. Dwell-based behaviors include the duration of dwell time in the image / text link display area, the browsing time on the product details page, and the dwell time in the core function introduction section—actions that reflect the user's level of attention to product information. Action-based behaviors include adding products to favorites or cart via image / text links, initiating inquiries, and canceling actions—actions with practical significance. These behaviors reflect users' interest in and willingness to purchase products. The classification process strictly follows preset rules, and each type of behavior is then quantified. Click-related behaviors are assigned corresponding quantitative values ​​based on the importance of the click, such as clicking the "add to cart" button having a higher quantitative value than simply clicking the details page. Dwell-related behaviors are converted into corresponding quantitative results based on the length of the dwell time; the longer the dwell time, the higher the user's attention, and the higher the quantitative value. Operation-related behaviors are assigned different quantitative values ​​based on the degree of completion of the operation and its impact on the purchase decision, such as completing the "add to cart" having a higher quantitative value than simply adding to favorites. The quantification process ensures that the intensity of user interaction and preference can be accurately quantified, transforming abstract behavioral feedback into statistically comparable quantitative results. The quantitative results of the three types of behaviors are integrated into dynamic behavioral supplementary data. This data comprehensively summarizes user feedback on recommended content, clearly presenting users' interests, focus, and behavioral tendencies towards recommended products, providing specific and valuable basis for data updates. The dynamic behavioral supplementary data is then fed back to the data collection and fusion steps to update the static feature quantitative data and dynamic behavior analysis data in a timely manner, such as... Figure 2 As shown, static feature quantification data can be integrated with the latest changes in users' consumption preferences, no longer limited to past historical data, while dynamic behavior analysis data can reflect the latest real-time needs of users, ensuring that the data always remains timely and accurate.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for analyzing and accurately recommending user behavior in live-streaming e-commerce, characterized in that, The specific steps of this method are as follows: S100. Data Acquisition and Fusion: Collect historical consumption data of users and perform static feature quantification to obtain static feature quantified data; Simultaneously collect and analyze the dynamic behavior data of users in the live broadcast room to obtain dynamic behavior analysis data; The static feature quantification data and the dynamic behavior analysis data are weighted and fused using a time-series weighted fusion algorithm to output a comprehensive user interest quantification value. S200, Comprehensive Parameter Generation: Based on the quantitative value of user comprehensive interest, analyze the online duration, exposure, sales trend and inventory status of each product in the live broadcast room, determine the product life cycle stage and determine the adaptation coefficient; use three-dimensional layered quantification technology to layer and quantify user interest in three dimensions: category, function and scenario, and generate three-dimensional quantitative values ​​of user interest; integrate the adaptation coefficient and the three-dimensional quantitative values ​​of user interest to form a comprehensive parameter set; S300, Recommendation scheme formulation: Receive a comprehensive parameter set, calculate the comprehensive fit degree using a multi-dimensional adaptation algorithm; recommend product display formats based on the comprehensive fit degree, customize scenario-based recommendation content, and integrate them to form a personalized recommendation scheme; S400, Recommendation Execution: Receive personalized recommendation schemes, and use the real-time recommendation engine in the live broadcast room to display recommended content to target users in predetermined interface positions in the product display area, bullet screen interaction bar, or side floating window, and collect interactive behavior data of target users in response to the recommended content in real time; S500, Closed-loop iteration: Receive the interactive behavior data of the target user, classify and quantify the interactive behavior data according to the preset classification rules, and generate dynamic behavior supplementary data; send the dynamic behavior supplementary data back to the data acquisition and fusion step to update the static feature quantification data and dynamic behavior analysis data.

2. The method for analyzing and accurately recommending live-streaming e-commerce user behavior according to claim 1, characterized in that, In step S100, the process of collecting users' historical consumption data and performing static feature quantification to obtain static feature quantified data is as follows: collecting users' historical consumption data on the live e-commerce platform, cleaning the collected historical consumption data, and obtaining standardized static data. Static feature quantization is performed on standardized static data to generate static feature quantized data. The process of synchronously collecting and parsing the dynamic behavior data of users in the live broadcast room to obtain dynamic behavior parsing data is as follows: synchronously collecting the dynamic behavior data of users in the live broadcast room, classifying the collected dynamic behavior data, and distinguishing between structured behavior data and unstructured behavior data. Quantify structured behavioral data and parse unstructured content; The parsed structured and unstructured behavioral data are integrated to generate dynamic behavioral parsing data.

3. The method for analyzing and accurately recommending live-streaming e-commerce user behavior according to claim 1, characterized in that, In step S100, the calculation formula of the time-weighted fusion algorithm is as follows: ,in, Quantify the user's overall interests; The static feature weight ratio coefficient; Quantify data for static features; This represents the total number of user dynamic behaviors collected within the live stream. As the basis for dynamic behavior weights; This is the time decay coefficient; For the first The time difference between the occurrence time of a dynamic behavior and the current calculation time; For the first Analyze dynamic behavior data.

4. The method for analyzing and accurately recommending live-streaming e-commerce user behavior according to claim 1, characterized in that, In step S200, the determination rule for determining the product lifecycle stage and matching coefficient in the live broadcast room is as follows: analyze the online time, exposure, sales trend and inventory status data of each product in the live broadcast room, divide the product lifecycle into preheating period, rising period, peak period and decline period, match it to the corresponding lifecycle stage, and configure a unique matching coefficient for each stage.

5. The method for analyzing and accurately recommending live-streaming e-commerce user behavior according to claim 1, characterized in that, In step S200, the specific steps for using three-dimensional layered quantification technology to layer and quantify user interests in three dimensions—category, function, and scenario—to generate three-dimensional quantified values ​​of user interests are as follows: combining the user's comprehensive interest quantified value with historical user consumption data and live stream dynamic behavior data to determine the user's category interest direction; based on the determined category interest direction, extracting the functional attributes of products clicked, added to cart, favorited, and purchased by the user within the category to extract the user's functional interest preferences; matching the real-time scenario of the live stream, and combining the user's past consumption behavior in the corresponding scenario with behavioral data within the live stream scenario to determine the user's scenario interest tendency; The three dimensions of category interest, function interest preference, and scenario interest tendency are quantified to generate a three-dimensional quantitative value of user interest that includes the quantification results of category, function, and scenario dimensions.

6. The method for analyzing and accurately recommending live-streaming e-commerce user behavior according to claim 1, characterized in that, In step S300, the calculation formula for the multi-dimensional adaptation algorithm is as follows: ,in, For overall compatibility; Product compatibility coefficient; This represents the weighting coefficient for the category interest dimension. Quantify user category interests; For the functional interest dimension weight coefficient; Quantify user interest in features; For the scene interest dimension weight coefficient; Quantify user scenario interests; Quantify the user's overall interests.

7. The method for analyzing and accurately recommending live-streaming e-commerce user behavior according to claim 1, characterized in that, In step S300, the process of integrating to form a personalized recommendation scheme is as follows: The system is divided into three intervals based on overall fit: high fit, medium fit, and low fit. The high fit interval's recommended product display area is a full-screen pop-up card; the medium fit interval's recommended interactive bar is a text and image link; and the low fit interval's recommended side floating window is a simple card. Based on the three-dimensional quantitative results of user interests associated with the corresponding interval, product attributes, and the real-time scene of the live stream, customized scenario-based recommendation content that highlights interest and scene fit is created. The recommended display format, contextualized recommended content, display duration, and display location are integrated to form a personalized recommendation scheme.

8. The method for analyzing and accurately recommending live-streaming e-commerce user behavior according to claim 1, characterized in that, In step S400, the real-time recommendation engine includes a data receiving module, a resource scheduling module, a display control module, and a behavior collection module, with the following specific structure: The data receiving module receives personalized recommendation schemes and parses the display format, content, location, and target user information; The resource scheduling module coordinates with various display interfaces in the live stream to allocate display resources. The display control module controls the loading, display time, and timing of content at predetermined locations based on the recommended scheme. The behavior collection module captures real-time data on target users' interactive behaviors such as clicking, staying, and closing recommended content.

9. The method for analyzing and accurately recommending live-streaming e-commerce user behavior according to claim 1, characterized in that, In step S500, the preset classification rules are constructed based on interactive behavior data; the interactive behavior data is classified and quantified according to the preset classification rules to generate dynamic behavior supplementary data. The specific process is as follows: the collected interactive behavior data is classified into click type, dwell type, and operation type and the classification is completed. Each type of behavior is quantified separately and transformed into statistically verifiable results. The quantification results of the three types of behaviors are then integrated into dynamic behavioral supplementary data.

Images