An interactive live broadcast system and method for network live broadcast marketing

Abstract

This invention discloses an interactive live streaming system and method for online live streaming marketing, relating to the field of online live streaming technology. It includes a live streaming acquisition module, a multimodal interactive data acquisition module, a user profile construction module, an interactive intent parsing module, a dynamic interactive strategy generation module, a live streaming content collaboration module, a data storage module, and a cloud service scheduling module. The output of the live streaming acquisition module is connected to the input of the cloud service scheduling module. The live streaming acquisition module is used to acquire real-time video streams, audio streams, and product display data from the broadcaster, and transmits the acquired data to the cloud service scheduling module. The output of the multimodal interactive data acquisition module is connected to the input of the cloud service scheduling module. This invention achieves personalized adaptation of interactive content and deep collaboration of live streaming content, improving audience engagement and marketing conversion efficiency, while enhancing the stability of the system under high-concurrency scenarios, making it suitable for various live streaming e-commerce marketing scenarios.

Description

Technical Field

[0001] This invention relates to the field of live streaming technology, and in particular to an interactive live streaming system and a method for live streaming marketing. Background Technology

[0002] With the development of the digital economy, live streaming marketing has become an important bridge connecting businesses and consumers. Interactivity is the core advantage of live streaming marketing, directly affecting audience engagement, retention rates, and final purchase conversion. Although existing interactive technologies for live streaming marketing have made some progress, they still have many shortcomings and cannot meet the needs of precise marketing and immersive interaction.

[0003] In existing technologies, most interactive live streaming systems adopt a traditional rule-driven interaction model, which pre-sets fixed interactive trigger conditions and response content. For example, a fixed reply is triggered when a viewer sends a specific keyword, or a preset animation is displayed when the gift donation reaches a certain amount. This model has obvious drawbacks: First, it lacks personalization and cannot provide differentiated interactive content based on different users' spending power and interests, resulting in insufficient interactive appeal. Second, it does not fully utilize multi-source interactive data and fails to effectively integrate multi-modal data such as bullet comments, voice, and behavior, making it difficult to accurately capture users' real needs. Third, the interactive strategy lacks dynamic optimization capabilities and cannot adjust the interactive logic in real time according to changes in traffic and product sales progress during the live stream, resulting in interaction delays in high-concurrency scenarios or failure to effectively reach potential consumers during key stages of product promotion.

[0004] Furthermore, existing interactive live streaming technologies that incorporate AI algorithms, such as semantic parsing technology based on natural language processing and interaction technology based on digital humans, often focus on single-dimensional interactive responses, failing to achieve deep integration of algorithms with marketing scenarios and thus unable to effectively convert interactive behavior into marketing results. For example, patent CN119166791A only achieves matching user questions with template answers, without combining product information and user consumption profiles for precise marketing recommendations; while patent CN120499405A, although achieving interactive-driven plot branching, does not consider the product conversion needs in marketing scenarios. Therefore, developing an interactive live streaming system and method for online live streaming marketing that can integrate multimodal data, accurately construct user profiles, and dynamically optimize interactive strategies is of significant practical importance. Summary of the Invention

[0005] This invention discloses an interactive live streaming system and method for online live streaming marketing. It constructs accurate user profiles through multimodal interactive data collection and a weighted Naive Bayes algorithm, and analyzes user intent using an attention mechanism-based multimodal data fusion algorithm. This achieves personalized adaptation of interactive content, providing differentiated interactive strategies based on different users' spending power and interests, significantly improving interactive appeal and user engagement. By employing a Q-learning-based reinforcement learning optimization algorithm with audience retention and product conversion rates as core objectives, it dynamically adjusts interactive strategies based on real-time data during the live stream, overcoming the rigidity and inflexibility of existing interactive strategies and improving the system's adaptability to changes in live streaming scenarios. The system also integrates personalized content through a live streaming content collaboration module. The interactive strategy is deeply integrated with the live video stream and product presentation rhythm, achieving an organic connection between interactive behavior and marketing content. This effectively guides audience interaction to the product understanding and purchase stage, significantly improving marketing conversion efficiency. Through a cloud service scheduling module, collaborative scheduling and data interaction among various modules are achieved, optimizing data transmission and processing flows. This effectively addresses the interactive data processing needs in high-concurrency live streaming scenarios, ensuring real-time interactive responses and system stability. By integrating multimodal interactive data such as text, voice, and behavior, it supports diverse interactive formats, such as personalized scripted responses, customized interactive activities, and precise product recommendations, enriching the interactive scenarios of live streaming marketing and enhancing the user's immersive experience. In summary, this solves the problems mentioned in the background technology.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0007] The present invention provides an interactive live streaming system for online live streaming marketing, comprising a live streaming acquisition module, a multimodal interactive data acquisition module, a user profile construction module, an interactive intent parsing module, a dynamic interactive strategy generation module, a live streaming content collaboration module, a data storage module, and a cloud service scheduling module.

[0008] The output end of the live streaming acquisition module is connected to the input end of the cloud service scheduling module. The live streaming acquisition module is used to acquire real-time video streams, audio streams, and product display screen data of the anchor, and transmit the acquired data to the cloud service scheduling module.

[0009] The output end of the multimodal interactive data acquisition module is connected to the input end of the cloud service scheduling module. The multimodal interactive data acquisition module is used to collect barrage text data, voice communication data, gift reward data, product click data and dwell time data from the audience end, and transmit the collected multi-source interactive data to the cloud service scheduling module.

[0010] The bidirectional port of the data storage module is connected to the cloud service scheduling module. The data storage module has a built-in user history database, product information database and interactive strategy template library. The data storage module is used to receive and store real-time data transmitted by the cloud service scheduling module, and at the same time provide historical data query services to the cloud service scheduling module.

[0011] The input end of the user profile building module is connected to the output end of the cloud service scheduling module, and the output end of the user profile building module is connected to the input end of the interaction intent parsing module. The user profile building module has a built-in user tag generation algorithm based on weighted Naive Bayes, which is used to receive real-time user interaction data transmitted by the cloud service scheduling module and user historical data retrieved by the data storage module. The user profile building module generates a precise user profile containing consumption capacity tags, interest preference tags, and interaction habit tags through the algorithm, and transmits the precise user profile to the interaction intent parsing module.

[0012] The input end of the interactive intent parsing module is also connected to the output end of the cloud service scheduling module, and the output end of the interactive intent parsing module is connected to the input end of the dynamic interaction strategy generation module. The interactive intent parsing module has a built-in multimodal data fusion algorithm based on the attention mechanism, which is used to receive real-time multimodal interactive data transmitted by the cloud service scheduling module and accurate user profiles transmitted by the user profile building module. The interactive intent parsing module generates intent tags containing consultation intent, purchase intent, and entertainment intent by parsing user interaction intent through the algorithm, and transmits the intent tags to the dynamic interaction strategy generation module.

[0013] The input end of the dynamic interaction strategy generation module is also connected to the output end of the cloud service scheduling module, and the output end of the dynamic interaction strategy generation module is connected to the input end of the live content collaboration module. The dynamic interaction strategy generation module has a built-in reinforcement learning optimization algorithm based on Q-learning, which is used to receive intent tags transmitted by the interaction intent parsing module, real-time live data transmitted by the cloud service scheduling module, and product information data retrieved by the data storage module. The dynamic interaction strategy generation module selects suitable basic interaction strategies from the interaction strategy template library through the algorithm and performs dynamic optimization. The dynamic interaction strategy generation module generates personalized interaction strategies that include interactive dialogue, interactive activities, and product recommendation schemes, and transmits the personalized interaction strategies to the live content collaboration module.

[0014] The input end of the live content collaboration module is also connected to the output end of the cloud service scheduling module. The output end of the live content collaboration module is connected to the input end of the cloud service scheduling module. It is used to receive personalized interaction strategies transmitted by the dynamic interaction strategy generation module and real-time live data transmitted by the cloud service scheduling module. It integrates personalized interaction strategies with live video stream and product explanation rhythm. The live content collaboration module generates collaborative interactive live stream and transmits it to the cloud service scheduling module, which then distributes it to the audience.

[0015] The cloud service scheduling module also maintains bidirectional data interaction with the live streaming acquisition module, multimodal interactive data acquisition module, user profile construction module, interactive intent parsing module, dynamic interactive strategy generation module, live streaming content collaboration module, and data storage module to realize data transmission and collaborative scheduling between the modules.

[0016] Furthermore, the multimodal interactive data acquisition module includes a text acquisition unit, a voice acquisition unit, a behavior acquisition unit, and a data preprocessing unit. The text acquisition unit is used to acquire audience bullet screen text data. The voice acquisition unit is used to acquire audience voice interaction data and perform voice-to-text processing. The behavior acquisition unit is used to acquire audience gift reward data, product click data, and dwell time data. The data preprocessing unit is used to perform noise reduction and standardization processing on the acquired multi-source data. The output end of the data preprocessing unit is connected to the input end of the cloud service scheduling module.

[0017] Furthermore, the user profile building module also includes a tag updating unit. The input end of the tag updating unit is connected to the output end of the cloud service scheduling module, and the output end of the tag updating unit is connected to the input end of the data storage module. The tag updating unit is used to dynamically update the user profile tags according to the user's real-time interaction data, and store the updated user profile in the data storage module.

[0018] Furthermore, the interactive intent parsing module also includes an intent verification unit. The input end of the intent verification unit is connected to the output end of the interactive intent parsing module, and the output end of the intent verification unit is connected to the input end of the dynamic interactive strategy generation module. The intent verification unit is used to verify the parsed interactive intent tags through historical interactive data and transmit the verified intent tags to the dynamic interactive strategy generation module.

[0019] Furthermore, the dynamic interactive strategy generation module also includes a strategy evaluation unit. The input end of the strategy evaluation unit is connected to the output end of the live content collaboration module, and the output end of the strategy evaluation unit is connected to the input end of the dynamic interactive strategy generation module. The strategy evaluation unit is used to evaluate the effect of the personalized interactive strategy based on the feedback data from the audience end, and transmit the generated evaluation results to the dynamic interactive strategy generation module to guide the optimization of the interactive strategy.

[0020] Furthermore, the live streaming content collaboration module includes a script integration unit, an activity embedding unit, a product recommendation unit, and a stream synthesis unit. The script integration unit is used to integrate interactive scripts with the anchor's audio stream. The activity embedding unit is used to embed interactive activity information into the live streaming video stream. The product recommendation unit is used to coordinate product recommendation schemes with product display screens. The stream synthesis unit is used to synthesize the integrated audio stream and video stream into a collaborative interactive live streaming stream.

[0021] An interactive live streaming method for online marketing includes the following steps:

[0022] Step S1: During the live broadcast start-up phase, the live broadcast acquisition module collects the host's real-time video stream, audio stream, and product display screen data. The multimodal interactive data acquisition module starts the audience's interactive data collection. At the same time, the cloud service scheduling module retrieves the user's historical data and product information data from the data storage module.

[0023] Step S2: User Profile Construction Stage. The user profile construction module obtains real-time and historical user interaction data through the cloud service scheduling module, calls the user tag generation algorithm based on weighted Naive Bayes to extract features and classify tags from user data, and generates accurate user profiles. The user tag generation algorithm based on weighted Naive Bayes assigns weight coefficients to data of different feature dimensions, where the weight coefficient for historical consumption amount is 0.35, the weight coefficient for real-time interaction frequency is 0.25, the weight coefficient for product click count is 0.2, and the weight coefficient for dwell time is 0.2. After weighted calculation, user tags are output.

[0024] Step S3: Interactive Intent Parsing Stage. The interactive intent parsing module obtains real-time multimodal interactive data and precise user profiles through the cloud service scheduling module. It calls the attention-based multimodal data fusion algorithm to perform feature fusion and intent recognition on multi-source interactive data such as text, voice, and behavior, generating user interactive intent tags. The attention-based multimodal data fusion algorithm evaluates the importance of different modal data through an attention weight allocation layer. The attention weight for voice chat data is 0.4, the attention weight for product click data is 0.3, the attention weight for bullet screen text data is 0.2, and the attention weight for gift reward data is 0.1. After attention-weighted fusion, the intent recognition result is output.

[0025] Step S4: Interactive Strategy Generation Stage. The dynamic interactive strategy generation module obtains user interaction intent tags, real-time live streaming data, and product information data through the cloud service scheduling module. It then calls a Q-learning-based reinforcement learning optimization algorithm to select basic interactive strategies from the interactive strategy template library and dynamically optimize them to generate personalized interactive strategies. The Q-learning-based reinforcement learning optimization algorithm uses audience retention rate and product click conversion rate as reward functions. It iteratively updates the Q-value matrix to optimize the interactive strategy parameters, thereby achieving dynamic adaptation of the interactive strategy.

[0026] Step S5: Live Stream Collaborative Output Stage. The live stream content collaboration module obtains personalized interactive strategies and real-time live stream data through the cloud service scheduling module. It then integrates interactive scripts, interactive activities, product recommendation schemes, and live stream video streams to generate a collaborative interactive live stream, which is then distributed to the audience through the cloud service scheduling module.

[0027] Step S6: During the live broadcast, repeat steps S2 to S5 to achieve dynamic interactive optimization based on real-time data until the end of the live broadcast. Then, use the cloud service scheduling module to store the interactive data, user profile data, and conversion data of this live broadcast into the data storage module.

[0028] Furthermore, in step S2, the accurate user profile generated by the user profile construction module also includes a user activity tag. The user activity tag is calculated by weighting the real-time interaction frequency and the historical interaction frequency, with weighting coefficients of 0.6 and 0.4, respectively.

[0029] Furthermore, in step S4, the reward function of the Q-learning-based reinforcement learning optimization algorithm is set as follows: Where α is the retention rate weighting coefficient, with a value ranging from 0.4 to 0.6, β is the conversion rate weighting coefficient, with a value ranging from 0.4 to 0.6, C1 is the real-time audience retention rate, and C2 is the real-time product click conversion rate.

[0030] Furthermore, in step S6, after the live stream ends, the data storage module also classifies and stores the interactive strategy effect data of this live stream. The effect data includes the retention rate, conversion rate and user feedback data corresponding to different interactive strategies, which are used for subsequent updates and optimizations of the interactive strategy template library.

[0031] The present invention has the following advantages over the prior art:

[0032] (1) This invention constructs accurate user profiles by collecting multimodal interactive data and using a weighted Naive Bayes algorithm, and analyzes user intent by combining an attention mechanism multimodal data fusion algorithm, thereby achieving personalized adaptation of interactive content. It can provide differentiated interactive strategies based on different users' spending power and interest preferences, significantly improving the attractiveness of interaction and user participation.

[0033] (2) This invention adopts a reinforcement learning optimization algorithm based on Q-learning, with audience retention rate and product conversion rate as the core objectives. It can dynamically adjust the interaction strategy according to real-time data during the live broadcast, overcome the defects of fixed and rigid interaction strategies in the prior art, and improve the system's adaptability to changes in the live broadcast scene.

[0034] (3) This invention integrates personalized interactive strategies with live video streams and product explanation rhythm through a live content collaboration module, realizing the organic connection between interactive behavior and marketing content, effectively guiding audience interaction to the product understanding and purchase stage, and significantly improving marketing conversion efficiency;

[0035] (4) The present invention realizes the collaborative scheduling and data interaction of each module through the cloud service scheduling module, optimizes the data transmission and processing process, can effectively cope with the interactive data processing needs in the high-concurrency scenario of live broadcast, and ensures the real-time performance of interactive response and the stability of system operation.

[0036] (5) This invention integrates multimodal interactive data such as text, voice, and behavior to support diverse forms of interaction, such as personalized word replies, customized interactive activities, and precise product recommendations, thereby enriching the interactive scenarios of live marketing and enhancing the immersive experience of users. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of the system module connection relationship of the present invention;

[0039] Figure 2 This is a schematic diagram showing the relationship between the method steps of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Reference Figures 1-2 An interactive live streaming system for online live streaming marketing includes a live streaming acquisition module, a multimodal interactive data acquisition module, a user profile building module, an interactive intent parsing module, a dynamic interactive strategy generation module, a live streaming content collaboration module, a data storage module, and a cloud service scheduling module.

[0042] The output end of the live streaming acquisition module is connected to the input end of the cloud service scheduling module. The live streaming acquisition module is used to collect real-time video streams, audio streams and product display screen data of the anchor, and transmit the collected data to the cloud service scheduling module.

[0043] The output end of the multimodal interactive data acquisition module is connected to the input end of the cloud service scheduling module. The multimodal interactive data acquisition module is used to collect barrage text data, voice communication data, gift reward data, product click data and dwell time data from the audience, and transmit the collected multi-source interactive data to the cloud service scheduling module.

[0044] The bidirectional port of the data storage module is connected to the cloud service scheduling module. The data storage module has a built-in user history database, product information database and interactive strategy template library. The data storage module is used to receive and store real-time data transmitted by the cloud service scheduling module, and at the same time provide historical data query services to the cloud service scheduling module.

[0045] The input end of the user profile building module is connected to the output end of the cloud service scheduling module, and the output end of the user profile building module is connected to the input end of the interaction intent parsing module. The user profile building module has a built-in user tag generation algorithm based on weighted Naive Bayes, which is used to receive real-time user interaction data transmitted by the cloud service scheduling module and user historical data retrieved by the data storage module. The user profile building module generates a precise user profile containing consumption capacity tags, interest preference tags, and interaction habit tags through the algorithm, and transmits the precise user profile to the interaction intent parsing module.

[0046] The input end of the interactive intent parsing module is also connected to the output end of the cloud service scheduling module, and the output end of the interactive intent parsing module is connected to the input end of the dynamic interaction strategy generation module. The interactive intent parsing module has a built-in multimodal data fusion algorithm based on the attention mechanism, which is used to receive real-time multimodal interactive data transmitted by the cloud service scheduling module and accurate user profiles transmitted by the user profile building module. The interactive intent parsing module uses the algorithm to parse user interaction intent to generate intent tags containing consultation intent, purchase intent, and entertainment intent, and transmits the intent tags to the dynamic interaction strategy generation module.

[0047] The input end of the dynamic interaction strategy generation module is also connected to the output end of the cloud service scheduling module, and the output end of the dynamic interaction strategy generation module is connected to the input end of the live content collaboration module. The dynamic interaction strategy generation module has a built-in reinforcement learning optimization algorithm based on Q-learning, which is used to receive intent tags transmitted by the interaction intent parsing module, real-time live data transmitted by the cloud service scheduling module, and product information data retrieved by the data storage module. The dynamic interaction strategy generation module selects suitable basic interaction strategies from the interaction strategy template library through the algorithm and performs dynamic optimization. The dynamic interaction strategy generation module generates personalized interaction strategies that include interactive dialogue, interactive activities, and product recommendation schemes, and transmits the personalized interaction strategies to the live content collaboration module.

[0048] The input end of the live content collaboration module is also connected to the output end of the cloud service scheduling module. The output end of the live content collaboration module is connected to the input end of the cloud service scheduling module to receive personalized interactive strategies transmitted by the dynamic interactive strategy generation module and real-time live data transmitted by the cloud service scheduling module. The personalized interactive strategies are integrated with the live video stream and the rhythm of product explanation. The live content collaboration module generates a collaborative interactive live stream and transmits it to the cloud service scheduling module, which then distributes it to the audience.

[0049] The cloud service scheduling module also maintains bidirectional data interaction with the live streaming acquisition module, multimodal interactive data acquisition module, user profile building module, interactive intent parsing module, dynamic interactive strategy generation module, live streaming content collaboration module, and data storage module to realize data transmission and collaborative scheduling between the modules.

[0050] The multimodal interactive data acquisition module includes a text acquisition unit, a voice acquisition unit, a behavior acquisition unit, and a data preprocessing unit. The text acquisition unit is used to collect audience bullet screen text data, the voice acquisition unit is used to collect audience voice interaction data and perform voice-to-text processing, the behavior acquisition unit is used to collect audience gift reward data, product click data, and dwell time data, and the data preprocessing unit is used to perform noise reduction and standardization processing on the collected multi-source data. The output end of the data preprocessing unit is connected to the input end of the cloud service scheduling module.

[0051] The user profile building module also includes a tag updating unit. The input of the tag updating unit is connected to the output of the cloud service scheduling module, and the output of the tag updating unit is connected to the input of the data storage module. The tag updating unit is used to dynamically update the user profile tags based on real-time user interaction data and store the updated user profile in the data storage module.

[0052] The interactive intent parsing module also includes an intent verification unit. The input of the intent verification unit is connected to the output of the interactive intent parsing module, and the output of the intent verification unit is connected to the input of the dynamic interactive strategy generation module. The intent verification unit is used to verify the parsed interactive intent tags using historical interactive data and transmit the verified intent tags to the dynamic interactive strategy generation module. The dynamic interactive strategy generation module also includes a strategy evaluation unit. The input of the strategy evaluation unit is connected to the output of the live content collaboration module, and the output of the strategy evaluation unit is connected to the input of the dynamic interactive strategy generation module. The strategy evaluation unit is used to evaluate the effectiveness of personalized interactive strategies based on audience feedback data and transmit the generated evaluation results to the dynamic interactive strategy generation module to guide the optimization of interactive strategies.

[0053] The live streaming content collaboration module includes a script integration unit, an activity embedding unit, a product recommendation unit, and a stream synthesis unit. The script integration unit integrates interactive scripts with the host's audio stream. The activity embedding unit embeds interactive activity information into the live streaming video stream. The product recommendation unit coordinates product recommendation schemes with product display screens. The stream synthesis unit synthesizes the integrated audio stream and video stream into a collaborative interactive live streaming stream.

[0054] An interactive live streaming method for online marketing includes the following steps:

[0055] Step S1: During the live broadcast start-up phase, the live broadcast acquisition module collects the host's real-time video stream, audio stream, and product display screen data. The multimodal interactive data acquisition module starts the audience's interactive data collection. At the same time, the cloud service scheduling module retrieves the user's historical data and product information data from the data storage module.

[0056] Step S2: User Profile Construction Stage. The user profile construction module obtains real-time and historical user interaction data through the cloud service scheduling module, calls the user tag generation algorithm based on weighted Naive Bayes to extract features and classify tags from user data, and generates accurate user profiles. The user tag generation algorithm based on weighted Naive Bayes assigns weight coefficients to data of different feature dimensions, where the weight coefficient for historical consumption amount is 0.35, the weight coefficient for real-time interaction frequency is 0.25, the weight coefficient for product click count is 0.2, and the weight coefficient for dwell time is 0.2. After weighted calculation, user tags are output.

[0057] The core formula of the algorithm is:

[0058]

[0059] Where: y represents the user tag category (e.g., moderate spending power, interest in oily skin and beauty products, etc.), and X represents the user feature vector. These correspond to historical spending amount, real-time interaction frequency, number of product clicks, and dwell time, respectively. The weight coefficients for each feature dimension ( =0.35, =0.25, =0.2, =0.2), P(y) is the prior probability of label y. Features The conditional probability under label y, where Z is the normalization factor; the user label is output after weighted calculation using the above formula.

[0060] Step S3: Interactive Intent Parsing Stage. The interactive intent parsing module obtains real-time multimodal interactive data and precise user profiles through the cloud service scheduling module. It calls the attention-based multimodal data fusion algorithm to perform feature fusion and intent recognition on multi-source interactive data such as text, voice, and behavior, generating user interactive intent tags. The attention-based multimodal data fusion algorithm evaluates the importance of different modal data through an attention weight allocation layer. The attention weight for voice chat data is 0.4, the attention weight for product click data is 0.3, the attention weight for bullet screen text data is 0.2, and the attention weight for gift reward data is 0.1. After attention-weighted fusion, the intent recognition result is output.

[0061] The core fusion formula is:

[0062]

[0063] Where: H is the fused feature vector, and m is the number of modal data types (m=4, corresponding to voice chat, product clicks, bullet screen text, and gift rewards respectively). The attention weights for the k-th modality data ( ), The feature vector of the k-th modality is used; the intent recognition result is output after attention-weighted fusion.

[0064] Step S4: Interactive Strategy Generation Stage. The dynamic interactive strategy generation module obtains user interaction intent tags, real-time live streaming data, and product information data through the cloud service scheduling module. It then calls a Q-learning-based reinforcement learning optimization algorithm to select basic interactive strategies from the interactive strategy template library and dynamically optimize them to generate personalized interactive strategies. The Q-learning-based reinforcement learning optimization algorithm uses audience retention rate and product click conversion rate as reward functions. It iteratively updates the Q-value matrix to optimize the interactive strategy parameters, thereby achieving dynamic adaptation of the interactive strategy.

[0065] Step S5: Live Stream Collaborative Output Stage. The live stream content collaboration module obtains personalized interactive strategies and real-time live stream data through the cloud service scheduling module. It then integrates interactive scripts, interactive activities, product recommendation schemes, and live stream video streams to generate a collaborative interactive live stream, which is then distributed to the audience through the cloud service scheduling module.

[0066] Step S6: During the live broadcast, repeat steps S2 to S5 to achieve dynamic interaction optimization based on real-time data until the end of the live broadcast. Then, use the cloud service scheduling module to store the interaction data, user profile data, and conversion data of this live broadcast into the data storage module.

[0067] In step S2, the accurate user profile generated by the user profile building module also includes user activity tags. The user activity tags are calculated by weighting the real-time interaction frequency and the historical interaction frequency, with weighting coefficients of 0.6 and 0.4, respectively.

[0068] In step S4, the reward function of the Q-learning-based reinforcement learning optimization algorithm is set as follows: Where α is the retention rate weighting coefficient, ranging from 0.4 to 0.6, β is the conversion rate weighting coefficient, ranging from 0.4 to 0.6, C1 is the real-time audience retention rate, and C2 is the real-time product click-through conversion rate; the algorithm's Q-value update formula is:

[0069]

[0070] Where s represents the current live stream state (including real-time traffic, product sales progress, user interaction activity, etc.), a represents the currently executed interaction strategy, Q(s,a) represents the action value of executing strategy a in state s, γ represents the discount factor (ranging from 0.8 to 0.95), r represents the immediate reward after executing strategy a (i.e., the R value mentioned above), s' represents the new state transitioned to after executing strategy a, and a' represents all possible interaction strategies in the new state s'. The maximum action value under the new state s';

[0071] In step S6, after the live stream ends, the data storage module also categorizes and stores the interactive strategy effect data of this live stream. The effect data includes retention rate, conversion rate and user feedback data corresponding to different interactive strategies, which will be used for subsequent updates and optimizations of the interactive strategy template library.

[0072] In the specific implementation process, a live streaming acquisition module is set up, using a high-definition camera to capture the host's video stream, a professional microphone to capture the audio stream, and a screen recording module to capture the product display screen. When the live streaming event officially starts and the streaming is completed, the multimodal interactive data acquisition module is immediately activated and enters a full-time data acquisition state. Each unit works collaboratively according to the preset division of labor: the text acquisition unit establishes real-time communication with the live streaming platform's bullet screen interface, and uses streaming data capture technology to capture all bullet screen information sent by the audience in real time. It not only records the text content of the bullet screen, but also synchronously captures the bullet screen sending time, the sending user ID, the live streaming timestamp corresponding to the bullet screen sending time, and other auxiliary information. For example, when the host is explaining a skin care product, the audience sends inquiries such as "Is this face cream suitable for dry skin?" or "Can sensitive skin use it?", which are all captured completely and associated with the live streaming scene tag "skin care product explanation time".

[0073] The voice acquisition unit collects the audience's voice communication content in real time through the live chat channel. At the same time, it starts the built-in offline speech-to-text engine (supporting noise reduction preprocessing) to complete real-time transcription, converting the audio interaction data into a standardized text format. It intelligently corrects dialects and colloquial expressions during the chat. For example, if an audience says, "Host, I have combination skin. Will using this in the summer cause breakouts?", it will be accurately transcribed into standard text and retain information such as chat duration and speaking time.

[0074] The behavior collection unit establishes data linkage with the front-end interactive page through the tracking technology, continuously tracking and accurately recording the core behavioral data of the audience. Among them, the gift reward data includes the type of reward, the amount of reward, the number of rewards, and the timing of the reward. The product-related behavior data covers key information such as the number of times the product link is clicked, the click time, the dwell time on a single product page, and whether it is added to the shopping cart. For example, when the host introduces a certain dress, the audience clicks the product link 3 times, stays for 1 minute and 20 seconds, and then rewards 1 "heart" gift. This behavioral data will be fully recorded and associated with the "dress explanation" scenario.

[0075] After receiving the three types of multi-source data—text, speech transcription, and behavioral data—the data preprocessing unit initiates a multi-step data cleaning process: First, invalid data is removed using a noise reduction filtering algorithm, such as meaningless garbled bullet comments like "asdfg123," repeated bullet comments, background noise transcribed into text during live streaming, and interference data such as accidental product clicks (staying time less than 2 seconds). Then, a Min-Max standardization algorithm is used to unify the format and value range of data of different dimensions and magnitudes. For example, gift reward amounts (0-5000 yuan) and dwell time (0-300 seconds) are standardized to the [0,1] range. Finally, the processed valid data is structured and encapsulated to generate standardized data units containing "user ID-data type-collection time-live streaming scene tag-data content." These units are then transmitted to the cloud service scheduling module via a high-bandwidth data channel, providing high-quality, high-reliability data input for subsequent user profile construction tag extraction and interaction intent analysis, ensuring the accuracy and efficiency of the entire data processing process from the source.

[0076] Among them, the tag update unit of the user profile building module receives real-time user interaction data transmitted by the cloud service scheduling module. When it detects new consumption behavior or changes in interaction preferences of users, it immediately updates the user's consumption ability tags, interest preference tags, etc. The updated user profile is synchronously stored in the data storage module to ensure that the user profile always matches the latest behavioral characteristics of users, providing accurate data support for the generation of subsequent personalized interaction strategies.

[0077] The intent verification unit of the interactive intent parsing module automatically retrieves the user's historical interaction data from the data storage module after receiving the parsed interactive intent tag. It compares and verifies the current intent tag with the historical interactive intent. If the verification finds that the current parsed purchase intent deviates from the user's historical high-frequency consultation intent, the intent parsing process is re-triggered until the verified intent tag is output, ensuring the accuracy of intent recognition results and avoiding the failure of interaction strategy adaptation due to intent misjudgment.

[0078] The strategy evaluation unit of the dynamic interactive strategy generation module receives real-time audience feedback data transmitted by the live content collaboration module, including the response rate of interactive scripts, the participation rate of interactive activities, and the click conversion rate of product recommendations. It quantifies and scores the effect of the current personalized interactive strategy through a preset evaluation model. If the score is lower than the preset threshold, the evaluation result is fed back to the dynamic interactive strategy generation module, which drives the module to adjust and optimize the interactive strategy parameters to improve the adaptability of the interactive strategy and the marketing conversion effect.

[0079] Among them, the script integration unit of the live content collaboration module will synchronously integrate the generated personalized interactive scripts with the real-time audio stream of the anchor to ensure that the timing of the interactive scripts is consistent with the rhythm of the anchor's explanation. The activity embedding unit will embed customized interactive activities (such as targeted coupon distribution and exclusive flash sale activities) into the live video stream in the form of pop-up windows. The product recommendation unit will push product detail cards in sync with the anchor's explanation of corresponding product categories based on user interest preference tags. The stream synthesis unit will synthesize the integrated audio stream, video stream and various interactive elements into a unified collaborative interactive live stream to achieve seamless connection between interactive content and live content.

[0080] Among them, the cloud service scheduling module dynamically allocates the data processing tasks of each module through a load balancing algorithm. When the live broadcast enters a high-concurrency period and the amount of audience interaction data surges, it automatically increases the resource priority of the multimodal interactive data acquisition module and the data storage module to ensure the real-time transmission and storage of interactive data. At the same time, it reduces duplicate data requests between modules through data caching technology, improves data interaction efficiency, and ensures the stability of system operation and the real-time nature of interactive response in high-concurrency scenarios.

[0081] Among them, the user history database of the data storage module classifies and archives the user interaction data and updated user profile data of this live broadcast, the product information database updates the product sales data and inventory status in real time, and the interaction strategy template library records the parameter configuration and effect data of each effective interaction strategy in this live broadcast, providing reusable templates and data references for the generation of interaction strategies for subsequent similar live broadcast activities, and helping the continuous optimization of the interaction strategy template library.

[0082] Among them, the user tag generation algorithm based on weighted Naive Bayes makes the core features such as historical consumption amount and real-time interaction frequency have a more significant impact on user tag generation through differentiated weight allocation, thereby improving the accuracy of user profiles and reducing the interference of secondary features. This ensures that the generated user tags can accurately reflect the user's consumption ability and interest preferences, providing a reliable basis for the generation of personalized interaction strategies.

[0083] Among them, the multimodal data fusion algorithm based on the attention mechanism strengthens the contribution of core interactive data to intent parsing by assigning differentiated attention weights to different modal data such as voice interaction and product clicks, improves the accuracy of user interaction intent recognition, and enables the system to more accurately capture users' core needs such as consultation and purchase, thus providing support for precise interaction.

[0084] The method is implemented as follows:

[0085] Step S1: When the live stream starts, the live stream acquisition module starts to collect the host's video stream, audio stream and beauty product display screen, the multimodal interactive data acquisition module starts to collect audience interaction data, and the cloud service scheduling module retrieves the user's historical beauty purchase data, past interaction data and beauty product information for this live stream from the data storage module.

[0086] Step S2: The user profile building module obtains real-time user interaction data (such as viewers sending bullet comments like "What foundation is suitable for oily skin?" or clicking on a foundation product link) and historical data (such as the user's past purchases of oily skin skincare products). It then generates user tags using a weighted Naive Bayes algorithm: moderate spending power, preference for oily skin beauty products, proactive inquiry behavior, and high activity level.

[0087] Step S3: The interaction intent parsing module obtains the above real-time interaction data and user profile, and uses the attention mechanism multimodal data fusion algorithm to parse the user intent as "consulting on purchase information related to foundation suitable for oily skin";

[0088] Step S4: The dynamic interaction strategy generation module obtains user intent tags, real-time live broadcast data (the current live broadcast is explaining the foundation category), and product information (this live broadcast has 3 foundations suitable for oily skin). Through Q-learning reinforcement learning optimization algorithm, it selects basic script templates from the interaction strategy template library and optimizes and generates personalized interaction strategies: for example, "Hello, oily-skinned baby! The foundation you clicked on is specially designed for oily skin, and it lasts for 8 hours without fading. If you order now, you will also receive a matching primer~ If you want to know the specific ingredients, you can type '1' and I will explain it to you in detail!" At the same time, it generates a "Giveaway for Oily Skin Foundation" activity plan;

[0089] Step S5: The live content collaboration module integrates the personalized interactive scripts with the host's audio stream, embeds the lottery information into the live video stream in the form of a pop-up window, and coordinates the recommended foundation product information with the product display screen to synthesize a collaborative interactive live stream, which is then distributed to the audience through the cloud service scheduling module.

[0090] Step S6: During the live stream, when the user sends "1", the system repeats steps S2 to S5 to generate a personalized interactive strategy for ingredient explanation; after the live stream ends, the system stores the interactive data, user profile data, click conversion rate of each product, and other data of this live stream into the data storage module for subsequent updates and optimizations of the beauty live stream interactive strategy template library.

[0091] It should be noted that all the products mentioned above are for illustrative purposes only and can be matched and changed according to the actual products.

[0092] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An interactive live streaming system for online live marketing, characterized in that, It includes a live streaming acquisition module, a multimodal interactive data acquisition module, a user profile building module, an interactive intent parsing module, a dynamic interactive strategy generation module, a live streaming content collaboration module, a data storage module, and a cloud service scheduling module; The output end of the live streaming acquisition module is connected to the input end of the cloud service scheduling module. The live streaming acquisition module is used to acquire real-time video streams, audio streams, and product display screen data of the anchor, and transmit the acquired data to the cloud service scheduling module. The output end of the multimodal interactive data acquisition module is connected to the input end of the cloud service scheduling module. The multimodal interactive data acquisition module is used to collect barrage text data, voice communication data, gift reward data, product click data and dwell time data from the audience end, and transmit the collected multi-source interactive data to the cloud service scheduling module. The bidirectional port of the data storage module is connected to the cloud service scheduling module. The data storage module has a built-in user history database, product information database and interactive strategy template library. The data storage module is used to receive and store real-time data transmitted by the cloud service scheduling module, and at the same time provide historical data query services to the cloud service scheduling module. The input end of the user profile building module is connected to the output end of the cloud service scheduling module, and the output end of the user profile building module is connected to the input end of the interaction intent parsing module. The user profile building module has a built-in user tag generation algorithm based on weighted Naive Bayes, which is used to receive real-time user interaction data transmitted by the cloud service scheduling module and user historical data retrieved by the data storage module. The user profile building module generates a precise user profile containing consumption capacity tags, interest preference tags, and interaction habit tags through the algorithm, and transmits the precise user profile to the interaction intent parsing module. The input end of the interactive intent parsing module is also connected to the output end of the cloud service scheduling module, and the output end of the interactive intent parsing module is connected to the input end of the dynamic interaction strategy generation module. The interactive intent parsing module has a built-in multimodal data fusion algorithm based on the attention mechanism, which is used to receive real-time multimodal interactive data transmitted by the cloud service scheduling module and accurate user profiles transmitted by the user profile building module. The interactive intent parsing module generates intent tags containing consultation intent, purchase intent, and entertainment intent by parsing user interaction intent through the algorithm, and transmits the intent tags to the dynamic interaction strategy generation module. The input end of the dynamic interaction strategy generation module is also connected to the output end of the cloud service scheduling module, and the output end of the dynamic interaction strategy generation module is connected to the input end of the live content collaboration module. The dynamic interaction strategy generation module has a built-in reinforcement learning optimization algorithm based on Q-learning, which is used to receive intent tags transmitted by the interaction intent parsing module, real-time live data transmitted by the cloud service scheduling module, and product information data retrieved by the data storage module. The dynamic interaction strategy generation module selects suitable basic interaction strategies from the interaction strategy template library through the algorithm and performs dynamic optimization. The dynamic interaction strategy generation module generates personalized interaction strategies that include interactive dialogue, interactive activities, and product recommendation schemes, and transmits the personalized interaction strategies to the live content collaboration module. The input end of the live content collaboration module is also connected to the output end of the cloud service scheduling module. The output end of the live content collaboration module is connected to the input end of the cloud service scheduling module. It is used to receive personalized interaction strategies transmitted by the dynamic interaction strategy generation module and real-time live data transmitted by the cloud service scheduling module. It integrates personalized interaction strategies with live video stream and product explanation rhythm. The live content collaboration module generates collaborative interactive live stream and transmits it to the cloud service scheduling module, which then distributes it to the audience. The cloud service scheduling module also maintains bidirectional data interaction with the live streaming acquisition module, multimodal interactive data acquisition module, user profile construction module, interactive intent parsing module, dynamic interactive strategy generation module, live streaming content collaboration module, and data storage module to realize data transmission and collaborative scheduling between the modules.

2. The interactive live streaming system for online live marketing according to claim 1, characterized in that, The multimodal interactive data acquisition module includes a text acquisition unit, a voice acquisition unit, a behavior acquisition unit, and a data preprocessing unit. The text acquisition unit is used to acquire audience bullet screen text data. The voice acquisition unit is used to acquire audience voice interaction data and perform voice-to-text conversion. The behavior acquisition unit is used to acquire audience gift reward data, product click data, and dwell time data. The data preprocessing unit is used to perform noise reduction and standardization processing on the acquired multi-source data. The output of the data preprocessing unit is connected to the input of the cloud service scheduling module.

3. The interactive live streaming system for online live marketing according to claim 1, characterized in that, The user profile building module also includes a tag updating unit. The input of the tag updating unit is connected to the output of the cloud service scheduling module, and the output of the tag updating unit is connected to the input of the data storage module. The tag updating unit is used to dynamically update user profile tags based on real-time user interaction data and store the updated user profile in the data storage module.

4. The interactive live streaming system for online live marketing according to claim 1, characterized in that, The interactive intent parsing module also includes an intent verification unit. The input end of the intent verification unit is connected to the output end of the interactive intent parsing module, and the output end of the intent verification unit is connected to the input end of the dynamic interactive strategy generation module. The intent verification unit is used to verify the parsed interactive intent tags through historical interactive data and transmit the verified intent tags to the dynamic interactive strategy generation module.

5. The interactive live streaming system for online live marketing according to claim 1, characterized in that, The dynamic interactive strategy generation module also includes a strategy evaluation unit. The input end of the strategy evaluation unit is connected to the output end of the live content collaboration module, and the output end of the strategy evaluation unit is connected to the input end of the dynamic interactive strategy generation module. The strategy evaluation unit is used to evaluate the effect of personalized interactive strategies based on audience feedback data and transmit the generated evaluation results to the dynamic interactive strategy generation module to guide the optimization of interactive strategies.

6. The interactive live streaming system for online live marketing according to claim 1, characterized in that, The live streaming content collaboration module includes a script integration unit, an activity embedding unit, a product recommendation unit, and a stream synthesis unit. The script integration unit is used to integrate interactive scripts with the host's audio stream. The activity embedding unit is used to embed interactive activity information into the live streaming video stream. The product recommendation unit is used to coordinate product recommendation schemes with product display screens. The stream synthesis unit is used to synthesize the integrated audio stream and video stream into a collaborative interactive live streaming stream.

7. An interactive live streaming method for online live streaming marketing, providing method support for an interactive live streaming system for online live streaming marketing as described in any one of claims 1-6, characterized in that, Includes the following steps: Step S1: During the live broadcast start-up phase, the live broadcast acquisition module collects the host's real-time video stream, audio stream, and product display screen data. The multimodal interactive data acquisition module starts the audience's interactive data collection. At the same time, the cloud service scheduling module retrieves the user's historical data and product information data from the data storage module. Step S2: User Profile Construction Stage. The user profile construction module obtains real-time and historical user interaction data through the cloud service scheduling module, calls the user tag generation algorithm based on weighted Naive Bayes to extract features and classify tags from user data, and generates accurate user profiles. The user tag generation algorithm based on weighted Naive Bayes assigns weight coefficients to data of different feature dimensions, where the weight coefficient for historical consumption amount is 0.35, the weight coefficient for real-time interaction frequency is 0.25, the weight coefficient for product click count is 0.2, and the weight coefficient for dwell time is 0.

2. After weighted calculation, user tags are output. Step S3: Interactive Intent Parsing Stage. The interactive intent parsing module obtains real-time multimodal interactive data and accurate user profiles through the cloud service scheduling module, calls the attention-based multimodal data fusion algorithm to perform feature fusion and intent recognition on multi-source interactive data such as text, voice, and behavior, and generates user interactive intent tags. The attention-based multimodal data fusion algorithm evaluates the importance of different modal data through an attention weight allocation layer. The attention weight for voice chat data is 0.4, the attention weight for product click data is 0.3, the attention weight for bullet screen text data is 0.2, and the attention weight for gift reward data is 0.

1. The intention recognition result is output after attention weighted fusion. Step S4: Interactive strategy generation stage. The dynamic interactive strategy generation module obtains user interaction intent tags, real-time live data and product information data through the cloud service scheduling module, calls the reinforcement learning optimization algorithm based on Q-learning, selects basic interactive strategies from the interactive strategy template library and performs dynamic optimization to generate personalized interactive strategies. The Q-learning-based reinforcement learning optimization algorithm uses audience retention rate and product click conversion rate as reward functions. It optimizes the interaction strategy parameters by iteratively updating the Q-value matrix, thereby achieving dynamic adaptation of the interaction strategy. Step S5: Live Stream Collaborative Output Stage. The live stream content collaboration module obtains personalized interactive strategies and real-time live stream data through the cloud service scheduling module. It then integrates interactive scripts, interactive activities, product recommendation schemes, and live stream video streams to generate a collaborative interactive live stream, which is then distributed to the audience through the cloud service scheduling module. Step S6: During the live broadcast, repeat steps S2 to S5 to achieve dynamic interactive optimization based on real-time data until the end of the live broadcast. Then, use the cloud service scheduling module to store the interactive data, user profile data, and conversion data of this live broadcast into the data storage module.

8. The interactive live streaming method for online live streaming marketing according to claim 7, characterized in that, In step S2, the accurate user profile generated by the user profile construction module also includes user activity tags. The user activity tags are calculated by weighting the real-time interaction frequency and the historical interaction frequency, with weighting coefficients of 0.6 and 0.4, respectively.

9. The interactive live streaming method for online live streaming marketing according to claim 7, characterized in that, In step S4, the reward function of the Q-learning-based reinforcement learning optimization algorithm is set as follows: Where α is the retention rate weighting coefficient, with a value ranging from 0.4 to 0.6, β is the conversion rate weighting coefficient, with a value ranging from 0.4 to 0.6, C1 is the real-time audience retention rate, and C2 is the real-time product click conversion rate.

10. The interactive live streaming method for online live streaming marketing according to claim 7, characterized in that, In step S6, after the live stream ends, the data storage module also classifies and stores the interactive strategy effect data of this live stream. The effect data includes the retention rate, conversion rate and user feedback data corresponding to different interactive strategies, which are used for subsequent updates and optimizations of the interactive strategy template library.

Images