Live room automation operation method, system and device using AIGC technology

Abstract

The application discloses a live broadcast room automatic operation method, system and equipment using AIGC technology, relates to the live broadcast operation technical field, and comprises a data acquisition unit, an AIGC operation unit, a resource scheduling unit and a terminal display unit. The AIGC operation data is monitored through the data acquisition unit, NLP sentiment analysis and user behavior modeling are carried out through the AIGC operation unit, the user value level is predicted in real time, the optimal operation recommendation strategy is generated, dynamic strategy adjustment is supported, data real-time analysis and decision are realized, and the AIGC-driven automatic content is optimized. Through the resource scheduling unit combined with the cost and benefit model, the automatic optimization configuration of hardware resources is realized, AIGC live broadcast display and multi-terminal interaction optimization are carried out through the terminal display unit, the rendering resolution of each terminal is dynamically adjusted, the user experience and cost control are considered, the service quality is ensured, and the operation cost is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of live broadcast operation, and particularly relates to a live broadcast room automatic operation method, system and equipment using AIGC technology. BACKGROUND

[0002] The AIGC technology live broadcast room refers to a virtual live broadcast room created by using artificial intelligence generated content (AIGC) technology. The virtual characters driven by artificial intelligence technology can simulate human voice, expression and action to perform activities such as live broadcast, reporting and performance. These virtual anchors can perform live broadcast for 24 hours without interruption, saving labor costs.

[0003] However, the traditional live broadcast room operation lacks real-time multi-dimensional data analysis capability, and it is difficult to accurately capture user dynamic needs. In addition, there is a defect of extensive resource allocation. Since the traditional recommendation strategy is based on static tags, it cannot dynamically match the real-time behavior of users with the adaptability of live broadcast content, such as click frequency, emotional tendency, etc., which may lead to data processing lag and lack of personalization. In addition, the hardware resource configuration is fixed and cannot be dynamically adjusted according to real-time cost, user value level, etc., which may cause unstable service for high-value users or resource waste.

[0004] In view of the above technical defects, a solution is proposed. SUMMARY

[0005] The purpose of the present application is to solve the problem that the traditional AIGC live broadcast room operation lacks real-time multi-dimensional data analysis capability and is difficult to accurately capture user dynamic needs, and there is a defect of extensive resource allocation.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0007] The live broadcast room automatic operation method using AIGC technology comprises the following steps:

[0008] Step 1: The data acquisition unit is used to monitor AIGC operation data: the AIGC operation data includes user portrait parameters, real-time behavior parameters, live broadcast content parameters and system resource parameters;

[0009] Step 2: The AIGC operation unit is used to analyze AIGC operation data: the AIGC operation unit includes a real-time processing module, a dynamic generation module and a recommendation distribution module;

[0010] The real-time processing module establishes a user behavior analysis model, combines the user portrait parameters, real-time behavior parameters and live broadcast content parameters, and predicts the user value level;

[0011] The text instruction and the voice instruction are fused by the dynamic generation module, so that the 3D virtual person behavior driving is carried out, and the AIGC generation quality is evaluated in multiple dimensions by monitoring the text language fluency, user interest relevance, historical comparison novelty and interactive emotion score, so that the optimal operation recommendation strategy is generated;

[0012] The recommendation distribution module outputs the recommendation priority of the live room goods according to the user value level and the optimal operation recommendation strategy;

[0013] Step three, the resource scheduling unit is used to balance the AIGC control resource allocation: through system resource parameter analysis live operation total cost, and with live room total income dynamic cooperation, output resource elasticity configuration scheme;

[0014] Step four, the terminal display unit is used for AIGC live display and multi-terminal interaction optimization: through resource elasticity configuration scheme to adaptively render resolution.

[0015] Further, the specific process of establishing a user behavior analysis model to predict the user value level is as follows:

[0016] The user portrait parameters, real-time behavior parameters and live content parameters are preprocessed synchronously, and the specific process is as follows:

[0017] The user portrait parameters include age, gender, device model, historical behavior, consumption ability and interest clustering; the historical behavior includes historical single price and historical purchase times; the consumption behavior includes consumption amount and consumption frequency;

[0018] Preprocess the user portrait parameters: obtain the consumption level index Clv and the category preference probability P(c) through the user portrait parameters, and then obtain the user portrait score F1 by combining the consumption level index Clv and the category preference probability P(c);

[0019] The real-time behavior parameters include the product page click frequency Cf, the live room stay time Ts, the behavior density index Bdi and the sentiment tendency value Snt;

[0020] Preprocess the real-time behavior parameters: obtain the purchase intention index PI through the real-time behavior parameters; obtain the emotional excitement index EH by combining the behavior density index Bdi and the sentiment tendency value Snt; obtain the real-time behavior score F2 by combining the purchase intention index PI and the emotional excitement index EH;

[0021] The live content parameters include product attributes, scene context, user matching degree and promotion sensitivity; the product attributes include product selling price, cost price, size and material;

[0022] Preprocessing live content parameters: Obtain user matching degree Um and promotion sensitivity Ps through product attributes and scene context; Obtain live content score F3 by combining user matching degree Um and promotion sensitivity Ps;

[0023] Then, the user portrait score F1, real-time behavior score F2 and live content score F3 are combined to generate the user value prediction index JZ;

[0024] Set the evaluation interval of the user value prediction index JZ, determine the user value level by interval comparison and mark it.

[0025] Furthermore, the specific process of preprocessing user portrait parameters is as follows:

[0026] User profile parameters include age, gender, device model, historical behavior, spending power, and interest clustering; historical behavior includes historical average order value and number of purchases; and spending behavior includes spending amount and frequency.

[0027] Mark the number of all users in the live broadcast room as N0, and calculate the average customer order value based on the historical customer order value of N0 users. and the standard deviation of average order value ; Mark any user as i, and mark the historical average order value of user i as , and thus the historical average order value of user i is calculated through Z-Score Standardize; add up the consumption amounts of N0 users to get the total consumption amount , and then calculate the average value to obtain the average consumption ; Mark the consumption frequency of user i as , and then obtain the consumption level index Clv;

[0028] By clustering users' historical purchases, we determine the categories of their products. We label the total number of product categories as Nm, any product category as c, the number of purchases by user i in category c as Nc, and the total number of purchases across all categories in the user's history as Ni. We then use Laplace smoothing to correct and calculate the probability P(c) of user i's preference for category c.

[0029] Then, by combining the consumption level index Clv and the category preference probability P(c), we can obtain the user portrait score F1.

[0030] Furthermore, the specific process of preprocessing real-time behavior parameters is as follows:

[0031] Real-time behavior parameters include product page click frequency Cf, live broadcast room stay time Ts, behavior density index Bdi and sentiment tendency value Snt;

[0032] The purchase intention index PI is obtained by weighted summation of the commodity page click frequency Cf, the live broadcast room stay time Ts, the behavior density index Bdi and the sentiment tendency value Snt.

[0033] The emotional excitement index EH is obtained by combining the change rate of the behavior density index Bdi and the sentiment tendency value Snt in the data collection period

[0034] The real-time behavior score F2 is obtained by combining the purchase intention index PI and the emotional excitement index EH.

[0035] Further, the specific process of preprocessing the live broadcast content parameters is as follows:

[0036] The live broadcast content parameters include product attributes, scene context, user matching degree and promotion sensitivity; the product attributes include product selling price, cost price, size and material;

[0037] The user matching degree Um is obtained by comparing the product price Pj of any product j with the historical average unit price The live broadcast scene countdown time Td is obtained by the scene context, the total number of user behavior indicators is marked as Nv, any user behavior indicator is marked as v, and the time-varying amplitude Av of the change difference of the user behavior indicator v before and after the live broadcast scene countdown time Td is marked.

[0038] The threshold value of the time-varying amplitude Av is set When the time-varying amplitude Av exceeds the threshold value , it is determined that the user behavior indicator v has a mutation of 1; thus, the number of mutations of Nv user behavior indicators is accumulated and marked as Na; and the promotion sensitivity Ps is obtained by the ratio of the number of mutations of user behavior indicators Na to the live broadcast scene countdown time Td.

[0039] The live broadcast content score F3 is obtained by combining the user matching degree Um and the promotion sensitivity Ps.

[0040] Further, the specific operation process of the dynamic generation module is as follows:

[0041] The dynamic instruction library of the 3D virtual person is set by the proportion of each user value level of all users in the live broadcast room, so as to call the preset text instructions and voice instructions in the dynamic instruction library.

[0042] The text feature vector Qw and the voice feature vector Qy are extracted respectively by the text instructions and the voice instructions.

[0043] The multi-modal fusion feature vector Gq is constructed by integrating the text feature vector Qw and the voice feature vector Qy.

[0044] ​The speech output and action sequence of the 3D virtual person are driven by the multi-modal fusion feature vector Gq.

[0045] Further, the specific process of multi-dimensional evaluation of AIGC generation quality is:

[0046] The speech text content output by the live broadcast is detected by the text language detection software, the number of syntax errors d1 and the coherence score d2 are obtained, and the fluency score D1 is obtained by weighted integration, so as to evaluate the text language fluency;

[0047] Through the historical data of user behavior indicators, the emotional excitement index EH and the category preference probability P(c) are monitored, and the interest score D2 is obtained by weighted integration, so as to evaluate the user interest relevance;

[0048] By comparing the speech text content generated by the current live broadcast with the speech text content generated by the historical live broadcast, the novelty score D3 is calculated, so as to evaluate the historical comparison novelty;

[0049] The average value and standard deviation are calculated through the historical data of the sentiment tendency value Snt, so as to obtain the sentiment integration coefficient d3 and the sentiment fluctuation coefficient d4 in turn, and the interaction emotion score D4 is obtained by weighting with the behavior density index Bdi;

[0050] The fluency score D1, the interest score D2, the novelty score D3 and the interaction emotion score D4 are weighted and summed to obtain the AIGC comprehensive quality index QD;

[0051] The AIGC comprehensive quality index QD, the fluency score D1, the interest score D2, the novelty score D3 and the interaction emotion score D4 are integrated to generate an AIGC quality evaluation vector, so as to multi-dimensionally evaluate the AIGC generation quality and generate an optimal operation recommendation strategy.

[0052] Further, the process of obtaining the resource elasticity configuration scheme is:

[0053] The service health score Rf is obtained by weighted integration of GPU utilization rate, response delay time and network bandwidth, and the service health score Rf is multiplied by the live broadcast service time Tf to obtain the operation service cost R0 of the live broadcast room; the total transaction income R1 of the live broadcast room is monitored in real time by the GMV model;

[0054] The difference between the operation service cost R0 of the live broadcast room and the total transaction income R1 is used to calculate the target function Pr of live broadcast operation profit, and the resource elasticity configuration scheme of the current live broadcast is obtained and output by maximizing the target function Pr of live broadcast operation profit.

[0055] The live room automation operation system using the AIGC technology comprises a data acquisition unit, an AIGC operation unit, a resource scheduling unit and a terminal display unit, wherein the AIGC operation unit comprises a real-time processing module, a dynamic generation module and a recommendation distribution module, and the system applies the live room automation operation method using the AIGC technology described above;

[0056] The data acquisition unit, the AIGC operation unit, the resource scheduling unit and the terminal display unit are in communication connection; the real-time processing module, the dynamic generation module and the recommendation distribution module are in communication connection;

[0057] The data acquisition unit is used for monitoring AIGC operation data: the data acquisition unit monitors and collects AIGC operation data through the AIGC operation unit and the resource scheduling unit;

[0058] The AIGC operation unit is used for analyzing AIGC operation data: wherein the real-time processing module is used for establishing a user behavior analysis model and predicting a user value level; the dynamic generation module is used for multi-modal fusion for 3D virtual human behavior driving, multi-dimensional evaluation of AIGC generation quality and generation of an optimal operation recommendation strategy; and the recommendation distribution module is used for outputting a recommendation priority of a live room commodity;

[0059] The resource scheduling unit is used for balancing AIGC control resource allocation;

[0060] The terminal display unit is used for AIGC live display and multi-terminal interaction optimization.

[0061] The live room automation operation device using the AIGC technology comprises a processor and a memory storing a computer program, and when the computer program is run by the processor, the above-mentioned live room automation operation method using the AIGC technology is executed.

[0062] As described above, due to the adoption of the above technical solutions, the beneficial effects of the present application are:

[0063] The present application monitors AIGC operation data through the data acquisition unit, performs NLP sentiment analysis and user behavior modeling through the AIGC operation unit, predicts a user value level in real time, realizes fusion processing of user portraits, real-time behaviors and live content, and generates an optimal operation recommendation strategy, thereby supporting dynamic strategy adjustment, realizing real-time data analysis and decision-making, and optimizing AIGC-driven automated content.

[0064] The application realizes automatic optimization configuration of hardware resources, balances AIGC control resource allocation, and dynamically adjusts the rendering resolution of each terminal by combining the cost and benefit model of the resource scheduling unit, giving consideration to user experience and cost control, guaranteeing service quality and reducing operating costs, and being suitable for intelligent upgrading of e-commerce live streaming room scenes. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 A step schematic diagram of the workflow of the application is shown;

[0066] Figure 2 A connection schematic diagram of the system modules of the application is shown. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application. EMBODIMENT

[0068] As shown in the figure, Figure 1-Figure 2 the live streaming room automatic operation system using the AIGC technology includes a data acquisition unit, an AIGC operation unit, a resource scheduling unit, and a terminal display unit, wherein the AIGC operation unit includes a real-time processing module, a dynamic generation module, and a recommendation distribution module;

[0069] The data acquisition unit, the AIGC operation unit, the resource scheduling unit, and the terminal display unit are in communication connection; the real-time processing module, the dynamic generation module, and the recommendation distribution module are in communication connection;

[0070] The data acquisition unit is used for monitoring AIGC operation data; the AIGC operation unit is used for analyzing AIGC operation data; the resource scheduling unit is used for balancing AIGC control resource allocation; and the terminal display unit is used for AIGC live streaming display and multi-terminal interaction optimization.

[0071] The working steps are as follows:

[0072] S1, the data acquisition unit monitors AIGC operation data: the AIGC operation data includes user portrait parameters, real-time behavior parameters, live streaming content parameters, and system resource parameters;

[0073] S1-1, the user portrait parameters include age, gender, device model, historical behavior, consumption ability, and interest clustering; the consumption grade and category preference are quantified through the user portrait parameters;

[0074] S1-2, the real-time behavior parameters include commodity page click frequency, live room stay duration, behavior density index, and sentiment tendency value; the behavior density index is obtained by weighting and synthesizing user behavior indexes in a unit time, the user behavior indexes include comment quantity and add-to-cart quantity; the sentiment tendency value is obtained by NLP sentiment analysis quantization; the purchase intention and emotional excitement level are identified through the real-time behavior parameters;

[0075] S1-3, the live content parameters include commodity attribute, scene context, user matching degree, and promotion sensitivity; the user matching degree is obtained by comparing the commodity price with the historical average order value of the user; the promotion sensitivity is obtained by detecting the user behavior mutation frequency during the countdown period of the live scene; the commodity adaptability and promotion response force are dynamically evaluated through the live content parameters;

[0076] S1-4, the system resource parameters include GPU utilization, response delay time, and network bandwidth; the service stability of high-value users is ensured through the system resource parameters.

[0077] S2, the AIGC operation unit analyzes AIGC operation data: the AIGC operation unit includes a real-time processing module, a dynamic generation module, and a recommendation distribution module;

[0078] S2-1, a user behavior analysis model is established through the real-time processing module, and the user portrait parameters, real-time behavior parameters, and live content parameters are combined to predict the user value level;

[0079] The specific process of establishing a user behavior analysis model to predict the user value level is as follows:

[0080] The user portrait parameters, real-time behavior parameters, and live content parameters are preprocessed synchronously;

[0081] S2-101, the specific process of preprocessing the user portrait parameters is as follows:

[0082] The user portrait parameters include age, gender, device model, historical behavior, consumption ability, and interest cluster;

[0083] The historical behavior includes historical average order value and historical purchase frequency; the consumption behavior includes consumption amount and consumption frequency;

[0084] The number of all users in the live room is marked as N0, the historical average order value is calculated through the historical average order value of N0 users and the historical average order standard deviation ; any user is marked as i, the historical average order value of user i is marked as , and the historical average order value of user i is standardized through Z-Score;

[0085] The total consumption amount is obtained by accumulating and summing the consumption amounts of the N users , and then the average value is calculated to obtain the average consumption amount ; the consumption frequency of user i is marked as ;

[0086] and then the consumption level index Clv is obtained: ; wherein , and are the weight coefficients of the user's single price, consumption amount and consumption frequency, and , and are obtained by calculating the experimental data and presetting, satisfying preset conditions; when the consumption level index Clv is higher, the consumption level of user i is higher;

[0087] The historical purchase goods of the user are classified by interest clustering, the number of all product categories is marked as Nm, any product category is marked as c, the historical purchase times of user i in category c is marked as Nc, and the total purchase times of all categories of user historical behavior is marked as Ni;

[0088] The category preference probability P(c) of user i for category c is calculated by Laplace smoothing correction: ; wherein is a smoothing parameter, and ;

[0089] The user portrait score F1 is obtained by combining the consumption level index Clv and the category preference probability P(c): ;

[0090] The specific process of preprocessing real-time behavior parameters is S2-102:

[0091] The real-time behavior parameters include the product page click frequency Cf, the live room stay time Ts, the behavior density index Bdi and the sentiment tendency value Snt.

[0092] The behavior density index Bdi is: ; wherein is the data collection period; Lc is the comment amount, and La is the add-to-cart amount. and are the weight coefficients of the comment amount and the add-to-cart amount, respectively, and the weight coefficients and are fitted by regression analysis after a large amount of historical data is measured and regressed.

[0093] Sentiment tendency value Snt: This is a sentiment analysis method that classifies user comments through NLP sentiment analysis. The classified sentiment is assigned a value according to the category and mapped to the interval [-1, 1], where -1 represents the most negative sentiment, 1 represents the most positive sentiment, and 0 represents neutral sentiment.

[0094] The purchase intention index PI is obtained by weighted summing the product page click frequency Cf, the live broadcast room stay time Ts, the behavior density index Bdi and the emotional tendency value Snt;

[0095] Through the behavior density index Bdi and sentiment tendency value Snt in the data collection period The emotional excitement index EH is obtained by combining the change rate of ;in, For the data collection cycle The difference between the behavioral density index Bdi before and after; For the data collection cycle The difference between the emotional tendency values ​​Snt before and after;

[0096] By combining the purchase intention index PI and the emotional excitement index EH, we can obtain the real-time behavior score F2: ;

[0097] S2-103, the specific process of pre-processing live content parameters is as follows:

[0098] Live content parameters include product attributes, scene context, user matching, and promotion sensitivity;

[0099] Product attributes include product selling price, cost price, size and material;

[0100] Through the product price Pj and historical customer unit price of any product j Compare and obtain the user matching degree Um: ; When the user matching degree Um is higher, it means that the product price is more consistent with the historical average order value;

[0101] Obtain the live scene countdown length Td through the scene context, mark the total number of user behavior indicators as Nv, mark any user behavior indicator as v, and mark the difference in the change of the user behavior indicator v before and after the live scene countdown length Td as the time-varying amplitude Av;

[0102] Set the threshold of time-varying amplitude Av , when the variation amplitude Av exceeds the threshold , then it is determined that the user behavior indicator v has produced 1 mutation; thus, the number of mutations of Nv user behavior indicators is accumulated and marked as Na;

[0103] Then, the promotion sensitivity Ps is obtained by the ratio of the number of mutations Na of the user behavior index and the countdown length Td of the live scene: ;

[0104] The live content score F3 is obtained by combining the user matching degree Um and the promotion sensitivity Ps: ;

[0105] S2-104, the user portrait score F1, the real-time behavior score F2 and the live content score F3 are combined to generate the user value prediction index JZ: ;

[0106] wherein, , and are the weight coefficients of the user portrait score F1, the real-time behavior score F2 and the live content score F3 respectively, and , and are all greater than 0, ; when the user portrait score F1, the real-time behavior score F2 and the live content score F3 are higher, the user value prediction index JZ is higher, so that the user value level is higher;

[0107] The evaluation interval of the user value prediction index JZ is set, the user value level is determined by interval comparison and marked, the user value level includes high, medium and low, and the users with high value level are focused on conversion.

[0108] S2-2, the text instructions and voice instructions are fused by the dynamic generation module, so as to drive the behavior of the 3D virtual person, and the text language fluency, user interest relevance, historical comparison novelty and interactive emotion score are monitored, so as to evaluate the AIGC generation quality in multiple dimensions, and generate the optimal operation recommendation strategy;

[0109] S2-201, the dynamic instruction library of the 3D virtual person is set by the proportion of each user value level of all users in the live room, so as to call the preset text instructions and voice instructions in the dynamic instruction library;

[0110] The text feature vector Qw and the voice feature vector Qy are extracted respectively through the text instructions and the voice instructions;

[0111] The text feature vector Qw and the voice feature vector Qy are integrated to construct the multi-modal fusion feature vector Gq;

[0112] The voice output and action sequence of the 3D virtual person are driven by the multi-modal fusion feature vector Gq, wherein the voice output includes the tone, speed and timbre of the voice; the action sequence includes the joint angle and body posture.

[0113] S2-202, monitor the text language fluency, user interest relevance, historical comparison novelty, and interactive emotional score, the specific process is as follows:

[0114] The text language detection software is used to detect the voice text content output by the live broadcast, obtain the number of syntax errors d1 and the coherence score d2, and obtain the fluency score D1 by weighted synthesis, so as to evaluate the text language fluency;

[0115] Through the historical data of user behavior indicators, the emotional excitement index EH and the category preference probability P(c) are monitored, and the interest score D2 is obtained by weighted synthesis, so as to evaluate the user interest relevance;

[0116] S2-202-2, by comparing the voice text content generated by the current live broadcast with the voice text content generated by the historical live broadcast, the novelty score D3 is calculated, so as to evaluate the historical comparison novelty;

[0117] Among them, the current content is marked as c, the set H is constructed by the historical content corresponding to m periods, and the historical content corresponding to any period is marked as hc, so that the similarity S(c, hc) between the current content c and the historical content hc of any period is calculated by using the text similarity algorithm, and then the novelty score D3 is obtained: ;

[0118] The average value and the standard deviation are calculated through the historical data of the sentiment tendency value Snt, so that the sentiment whole system coefficient d3 and the sentiment fluctuation coefficient d4 are obtained in turn, and the behavior density index Bdi is weighted, so that the interactive emotional score D4 is obtained by comprehensive calculation;

[0119] S2-203, the fluency score D1, the interest score D2, the novelty score D3 and the interactive emotional score D4 are weighted and summed to obtain the AIGC comprehensive quality index QD;

[0120] The AIGC comprehensive quality index QD, the fluency score D1, the interest score D2, the novelty score D3 and the interactive emotional score D4 are integrated to generate an AIGC quality evaluation vector;

[0121] S2-204, the AIGC comprehensive quality index QD and the specific performance of each dimension are used to evaluate the AIGC generation quality in multiple dimensions, and the optimal operation recommendation strategy is generated;

[0122] For example, when the text language fluency is low, the natural language processing model is recommended to be optimized; when the user interest relevance is insufficient, it is suggested to adjust the content personalization combined with the user portrait; when the historical comparison novelty is low, it is encouraged to increase innovative elements; when the interactive emotional score is low, the emotional expression and interactive design of the content are optimized;

[0123] By establishing a strategy library and a rule engine, a corresponding recommended strategy is automatically matched according to different evaluation results, so as to improve the quality and operation effect of AIGC generated content.

[0124] S2-3, the recommended distribution module outputs the recommended priority of the live room goods according to the user value level and the optimal operation recommended strategy;

[0125] For example, for the case where high-value users account for a high proportion, high-margin new products are preferentially recommended; for the case where medium-value users account for a high proportion, blockbuster repeat purchase goods and potential new products are preferentially recommended; for the case where low-value users account for a high proportion, lead-in goods and promotion goods are preferentially recommended;

[0126] Thus, the live behavior of the 3D virtual person is dynamically adjusted, including live operation language and action expression state.

[0127] S3, the resource scheduling unit balances AIGC control resource allocation: through system resource parameter analysis live operation total cost, and dynamically cooperates with live room total transaction income, outputs resource elasticity configuration scheme;

[0128] S3-1, through weighted comprehensive acquisition service health score Rf by GPU utilization rate, response delay time, network bandwidth, and multiplying service health score Rf with live service time Tf, operation service cost R0 of the live room is acquired;

[0129] S3-2, through the GMV model, the total transaction income R1 of the live room is monitored in real time; wherein, the GMV model refers to the pricing or income calculation method based on the total transaction amount (Gross Merchandise Value) in e-commerce;

[0130] S3-3, through the difference between the operation service cost R0 of the live room and the total transaction income R1, the target function Pr of live operation profit is calculated, and the target function Pr of live operation profit is maximized, so as to output the resource elasticity configuration scheme of the current live;

[0131] The resource elasticity configuration scheme includes dynamically adjusting resource configuration of GPU utilization rate, response delay time, network bandwidth and multi-terminal interaction optimization operation, so as to make the live operation more efficient.

[0132] S4, the terminal display unit performs AIGC live display and multi-terminal interaction optimization: through the resource elasticity configuration scheme, the resolution is adaptively rendered, finally improving the new user conversion rate, reducing the live comprehensive cost, and improving the live room automation operation efficiency;

[0133] For example, when GPU computing power and bandwidth are insufficient, the AIGC resolution of the display port is reduced, thereby reducing the total cost of live operation.

[0134] The live room automatic operation equipment using the AIGC technology comprises a processor and a memory storing a computer program, and when the computer program is executed by the processor, the live room automatic operation method using the AIGC technology is executed.

[0135] In summary, the application monitors AIGC operation data through a data acquisition unit, performs NLP sentiment analysis and user behavior modeling through an AIGC operation unit, and predicts user value levels in real time, realizes the fusion processing of user portraits, real-time behaviors and live content, and generates optimal operation recommendation strategies, thereby supporting dynamic strategy adjustment, realizing real-time data analysis and decision-making, and optimizing AIGC-driven automatic content.

[0136] The application realizes automatic optimization of hardware resources through a resource scheduling unit combined with a cost and benefit model, balances AIGC control resource allocation, and realizes AIGC live display and multi-terminal interaction optimization through a terminal display unit, thereby dynamically adjusting the rendering resolution of each terminal, taking into account user experience and cost control, ensuring service quality and reducing operation cost, and being suitable for intelligent upgrading of e-commerce live room scenes.

[0137] The above embodiments can be realized by software, hardware, firmware or any combination thereof, in whole or in part. When realized by software, the above embodiments can be realized in the form of a computer program product in whole or in part. Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized by hardware or software depends on the specific application and design constraints of the technical solutions.

[0138] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, i.e. they can be located in one place or distributed on multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiment according to actual needs.

[0139] The above is only a preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, and any skilled person in the art can make equivalent replacements or changes to the technical solutions and inventive concepts of the application within the technical scope disclosed by the application, which should be covered by the protection scope of the application.

Claims

1. The automated operation method of a live broadcast room using AIGC technology is characterized by: The following steps are involved: Step 1: The data collection unit is used to monitor AIGC operation data: AIGC operation data includes user profile parameters, real-time behavior parameters, live content parameters, and system resource parameters; Step 2: The AIGC operation unit is used to analyze AIGC operation data: the AIGC operation unit includes a real-time processing module, a dynamic generation module, and a recommendation and distribution module; A user behavior analysis model is established through the real-time processing module, combining user portrait parameters, real-time behavior parameters, and live content parameters to predict user value levels; Through the dynamic generation module, multimodal fusion of text and voice commands is performed to drive 3D virtual human behavior. By monitoring the fluency of text language, relevance to user interests, historical comparison novelty, and interactive sentiment scores, the AIGC generation quality is evaluated in multiple dimensions to generate the optimal operation recommendation strategy. The recommendation distribution module outputs the recommendation priority of live broadcast room products based on user value level and optimal operation recommendation strategy; Step 3: The resource scheduling unit is used to balance the allocation of AIGC control resources: it analyzes the total cost of live broadcast operations through system resource parameters, dynamically coordinates with the total transaction revenue of the live broadcast room, and outputs a resource elasticity configuration plan; Step 4: The terminal display unit is used for AIGC live broadcast and multi-terminal interaction optimization: adaptive rendering resolution is achieved through resource elastic configuration scheme; The specific process of establishing a user behavior analysis model to predict user value levels is as follows: Synchronously pre-process the user portrait parameters, real-time behavior parameters, and live content parameters. The specific process is as follows: User profile parameters include age, gender, device model, historical behavior, spending power, and interest clustering; historical behavior includes historical average order value and number of purchases; and spending behavior includes spending amount and frequency. Preprocessing user profile parameters: Obtain the consumption level index Clv and category preference probability P(c) through the user profile parameters. Then, combine the consumption level index Clv and category preference probability P(c) to obtain the user profile score F1. Real-time behavior parameters include product page click frequency Cf, live broadcast room stay time Ts, behavior density index Bdi and sentiment tendency value Snt; Preprocessing real-time behavior parameters: Obtain the purchase intention index PI through real-time behavior parameters; obtain the emotional excitement index EH by combining the behavior density index Bdi and the emotional tendency value Snt; obtain the real-time behavior score F2 by combining the purchase intention index PI and the emotional excitement index EH; Live content parameters include product attributes, scene context, user matching, and promotion sensitivity; product attributes include selling price, cost price, size, and material; Preprocessing live content parameters: Obtain user matching degree Um and promotion sensitivity Ps through product attributes and scene context; Obtain live content score F3 by combining user matching degree Um and promotion sensitivity Ps; Then, the user portrait score F1, real-time behavior score F2 and live content score F3 are combined to generate the user value prediction index JZ; Set the evaluation interval of the user value prediction index JZ, determine the user value level by interval comparison and mark it.

2. The method for automated operation of a live broadcast room using AIGC technology according to claim 1, characterized in that: The specific process of preprocessing user portrait parameters is as follows: Mark the number of all users in the live broadcast room as N0, and calculate the average customer order value based on the historical customer order value of N0 users. and standard deviation of average order value ; Mark any user as i, and mark the historical average order value of user i as , and thus the historical average order value of user i is calculated through Z-Score Standardize; add up the consumption amounts of N0 users to get the total consumption amount , and then calculate the average value to obtain the average consumption ; Mark the consumption frequency of user i as , and then obtain the consumption level index Clv; By clustering users' historical purchases, we determine the categories of their products. We label the total number of product categories as Nm, any product category as c, the number of purchases by user i in category c as Nc, and the total number of purchases across all categories in the user's history as Ni. We then use Laplace smoothing to correct and calculate the probability P(c) of user i's preference for category c. Then, by combining the consumption level index Clv and the category preference probability P(c), we can obtain the user portrait score F1.

3. The method for automated operation of a live broadcast room using AIGC technology according to claim 2, characterized in that: The specific process of preprocessing real-time behavior parameters is as follows: Real-time behavior parameters include product page click frequency Cf, live broadcast room stay time Ts, behavior density index Bdi and sentiment tendency value Snt; The purchase intention index PI is obtained by weighted summing the product page click frequency Cf, the live broadcast room stay time Ts, the behavior density index Bdi and the emotional tendency value Snt; Through the behavior density index Bdi and sentiment tendency value Snt in the data collection period The emotional excitement index EH is obtained by combining the change rate of The real-time behavior score F2 is obtained by combining the purchase intention index PI and the emotional excitement index EH.

4. The method for automated operation of a live broadcast room using AIGC technology according to claim 3, characterized in that: The specific process of preprocessing live content parameters is as follows: Live content parameters include product attributes, scene context, user matching, and promotion sensitivity; product attributes include selling price, cost price, size, and material; Through the product price Pj and historical customer unit price of any product j Perform comparison and obtain user matching degree Um; Obtain the live scene countdown length Td through the scene context, mark the total number of user behavior indicators as Nv, mark any user behavior indicator as v, and mark the difference in the change of the user behavior indicator v before and after the live scene countdown length Td as the time-varying amplitude Av; Set the threshold of time-varying amplitude Av , when the variation amplitude Av exceeds the threshold , then it is determined that the user behavior indicator v has generated 1 mutation; thus, the number of mutations of Nv user behavior indicators is accumulated and marked as Na; then the promotion sensitivity Ps is obtained by the ratio of the number of mutations of the user behavior indicator Na to the countdown length Td of the live broadcast scene; By combining the user matching degree Um and the promotion sensitivity Ps, the live content score F3 is obtained.

5. The method for automated operation of a live broadcast room using AIGC technology according to claim 4, characterized in that: The dynamic generation module performs multimodal fusion of text commands and voice commands to drive 3D virtual human behavior. The specific operation process is as follows: By using the proportion of each user value level of all users in the live broadcast room, a dynamic instruction library of the 3D virtual person is set up, thereby calling the preset text instructions and voice instructions in the dynamic instruction library; Through text instructions and voice instructions, the text feature vector Qw and the voice feature vector Qy are extracted respectively; By integrating the text feature vector Qw and the speech feature vector Qy, a multimodal fusion feature vector Gq is constructed; The speech output and action sequence of the 3D virtual human are driven by the multimodal fusion feature vector Gq.

6. The method for automated operation of a live broadcast room using AIGC technology according to claim 5, characterized in that: The specific process of multi-dimensional evaluation of AIGC generation quality is as follows: The text content of the live broadcast output is tested using text language detection software to obtain the number of grammatical errors d1 and the coherence score d2, and then weighted and comprehensively calculated to obtain the fluency score D1, thereby evaluating the fluency of the text language; By monitoring the emotional excitement index (EH) and category preference probability (P(c)) through historical data of user behavior indicators, and performing weighted comprehensive analysis to obtain the interest score (D2), we can evaluate the relevance of user interests. By comparing the speech text content generated by the current live broadcast with the speech text content generated by the historical live broadcast, the novelty score D3 is calculated to evaluate the historical comparison novelty; The average value and standard deviation of the historical data of the emotional tendency value Snt are calculated to obtain the overall emotional coefficient d3 and the emotional fluctuation coefficient d4 in turn, and then weighted with the behavior density index Bdi to obtain the interactive emotional score D4; The AIGC comprehensive quality index QD is obtained by weighted summing the fluency score D1, interest score D2, novelty score D3 and interactive emotion score D4; By integrating the AIGC comprehensive quality index QD, fluency score D1, interest score D2, novelty score D3 and interactive emotion score D4, an AIGC quality evaluation vector is generated, thereby evaluating the AIGC generation quality in multiple dimensions and generating the optimal operation recommendation strategy.

7. The method for automated operation of a live broadcast room using AIGC technology according to claim 6, characterized in that: The process of obtaining a resource elastic configuration plan is as follows: The service health score Rf is obtained by weighting the GPU utilization, response delay time, and network bandwidth, and the service health score Rf is multiplied by the live service time Tf to obtain the operating service cost R0 of the live broadcast room; Monitor the total transaction revenue R1 of the live broadcast room in real time through the GMV model; The objective function Pr of the live broadcast operation profit is calculated by the difference between the operating service cost R0 of the live broadcast room and the total transaction revenue R1. By maximizing the objective function Pr of the live broadcast operation profit, the resource elasticity configuration plan of the current live broadcast is obtained and output.

8. The automated live broadcast room operation system using AIGC technology is characterized by: The system comprises a data collection unit, an AIGC operation unit, a resource scheduling unit and a terminal display unit, wherein the AIGC operation unit comprises a real-time processing module, a dynamic generation module and a recommendation distribution module. The system applies the live broadcast room automated operation method using AIGC technology as described in any one of claims 1 to 7 above; Communication connection between the data collection unit, AIGC operation unit, resource scheduling unit and terminal display unit; communication connection between the real-time processing module, dynamic generation module and recommendation distribution module; The data collection unit is used to monitor AIGC operation data: the data collection unit monitors and collects AIGC operation data through the AIGC operation unit and resource scheduling unit; The AIGC operation unit is used to analyze AIGC operation data: the real-time processing module is used to establish a user behavior analysis model and predict user value levels; the dynamic generation module is used to drive 3D virtual human behavior through multimodal fusion, and to evaluate the quality of AIGC generation in multiple dimensions to generate the optimal operation recommendation strategy; the recommendation distribution module is used to output the recommendation priority of products in the live broadcast room; The resource scheduling unit is used to balance the allocation of AIGC control resources; The terminal display unit is used for AIGC live broadcast and multi-terminal interaction optimization.

9. Live broadcast room automated operation equipment using AIGC technology, characterized by: It includes a processor and a memory storing a computer program. When the computer program is run by the processor, it executes the live broadcast room automatic operation method using AIGC technology as described in any one of claims 1 to 7.

Images