Internet live broadcast content recommendation method and system based on artificial intelligence

Abstract

The invention relates to an Internet live broadcast content recommendation method based on artificial intelligence, a comprehensive recommendation index is generated by acquiring user behaviors, live broadcast content characteristics and context environment parameters in real time and utilizing a dynamic weighting calculation model, and the method comprises the following steps: a content classification model scores live broadcast content; the personalized recommendation model analyzes the user recommendation index and the content score to generate an initial personalized content list; the user interaction analysis model generates recommendation adjustment parameters according to the user interaction data, and then revises recommendation indexes; and the content adjustment model optimizes the personalized content list according to the adjustment parameters, and recommends the live broadcast content to the user. The method has the effects of providing accurate and personalized live broadcast recommendation, enhancing user experience and improving platform operation efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of live broadcast push, and in particular to an artificial intelligence-based method and system for recommending Internet live broadcast content. Background Art

[0002] With the rapid development of internet technology, the live streaming industry has become a highly influential form of content dissemination and social interaction, experiencing explosive growth. However, in this vast landscape of live streaming content, users often face a dilemma in choosing the best option. The key challenge facing the industry is how to quickly deliver live streams that suit users' preferences.

[0003] Most existing live content recommendation systems rely on users' historical viewing records or simple interaction data. For example, they push the same type or the same live content based on the user's last viewing history. This makes it difficult to meet users' real-time needs. Users need to conduct precise searches or select pushed content multiple times to get the content they expect, which reduces the user's live viewing experience. Therefore, improvement is needed. Summary of the Invention

[0004] In order to improve users' live video viewing experience, this application provides an artificial intelligence-based Internet live content recommendation method and system.

[0005] In the first aspect, the above-mentioned invention object of the present application is achieved through the following technical solutions:

[0006] An artificial intelligence-based method for recommending live internet content, comprising the steps of:

[0007] Real-time acquisition of user behavior parameters, live content feature parameters, and context parameters, including:

[0008] User behavior parameters include user stay time conversion rate, cross-live room migration frequency, and real-time interaction heat value;

[0009] The live content feature parameters include the host’s voiceprint feature vector, picture color dynamic entropy, and topic tag density;

[0010] The context parameters include the network delay compensation coefficient and the period traffic attenuation factor;

[0011] The preset content dynamic weighted calculation model performs a fusion calculation on the user behavior parameters, live content feature parameters and context environment parameters, and generates a comprehensive recommendation index;

[0012] Acquire real-time live content data, and use a pre-set content classification model to perform a comprehensive content scoring on the live content data based on a machine self-learning algorithm;

[0013] The preset personalized recommendation generation model analyzes the user's comprehensive recommendation index and the comprehensive score of the live broadcast content based on a machine self-learning algorithm to generate an initial personalized content list for the user;

[0014] Acquire user interaction data when the user watches the live broadcast, and associate the user interaction data with the watched live broadcast content based on a preset common timeline to construct a user interaction dataset;

[0015] A preset user interaction analysis model analyzes the user interaction dataset based on a machine self-learning algorithm to generate recommended adjustment parameters, wherein the recommended adjustment parameters are used to modify the comprehensive recommendation index to generate a modified comprehensive recommendation index;

[0016] The preset content adjustment model adjusts the initial personalized content list based on the recommendation adjustment parameter to generate a revised personalized content list, and recommends live content to the user based on the revised personalized content list.

[0017] By adopting the above technical solution, the present invention obtains user behavior parameters, live content feature parameters and context environment parameters in real time, and generates a comprehensive recommendation index in combination with a dynamic weighted calculation model, thereby achieving deep integration of multi-dimensional data, which can more comprehensively reflect user needs and the actual situation of live content, thereby generating more accurate recommendation results. Most existing recommendation systems lack a dynamic adjustment mechanism and cannot be optimized according to real-time feedback from users. The present invention introduces user interaction data sets and recommendation adjustment parameters to be able to correct the initial recommendation index in real time, further optimize the recommended content, and make it more in line with the user's real-time preferences.

[0018] In a preferred example, the present application can be further configured as follows: in the step of performing a fusion calculation of the user behavior parameters, the live content feature parameters, and the context environment parameters by a preset content dynamic weighted calculation model to generate a comprehensive recommendation index, the content dynamic weighted calculation model is preset with a comprehensive recommendation index calculation formula as follows:

[0019] ,

[0020] + + =1;

[0021] in, The user behavior parameter weight is used to measure the importance of the user's own preferences. The weight of content feature parameters is used to measure the criticality of content quality. The context parameter weight is used to measure the adaptability of the real-time scene. is the information entropy parameter, is the normalized user parameter, is the dynamic correlation coefficient, is the context parameter, is the time decay factor, t is the time parameter, i is the i-th primary candidate content item, j is the j-th secondary candidate content item, To standardize content parameters.

[0022] By adopting the above technical solution, the comprehensive recommendation index calculation formula realizes dynamic weighted calculation of multi-dimensional data by integrating user behavior parameters, live content feature parameters and context environment parameters, thereby significantly improving the relevance and accuracy of the recommendation system. This comprehensive consideration can not only more accurately capture the user's current preferences and needs and improve the relevance of recommended content, but also significantly enhance the user experience, making users feel more intimate and satisfactory service. The dynamic weighted calculation model can be adjusted according to the user's real-time behavior and preference changes, so that the recommended content is more in line with the user's personalized needs, while improving the platform's operational efficiency, and increasing the exposure and viewing rate of live content through accurate recommendations. In addition, the introduction of information entropy parameters and dynamic correlation coefficients promotes content diversification, avoids the homogenization of recommended content, and helps users discover more interesting and valuable content. The consideration of time decay factor and contextual environment parameters enhances the adaptability of the recommendation system, enabling it to maintain good performance in different network conditions, time periods and user groups. The automated recommendation algorithm reduces manual intervention and operating costs, improves user satisfaction, reduces user churn, and thus reduces the marketing cost of acquiring new users. Real-time monitoring and analysis of user behavior data enhances the security of the platform. At the same time, the dynamic weighted calculation model can be adjusted according to the latest laws, regulations and policy requirements to ensure the compliance of recommended content. Accurate recommendation algorithms provide the platform with more business opportunities, such as advertising, content cooperation, etc., promote the platform's technological innovation and development, and improve the platform's competitiveness. In summary, through its innovative design, the formula of this solution has brought comprehensive performance improvements to the recommendation system of the live broadcast industry and provided users with a richer and more personalized viewing experience.

[0023] In a preferred example, the present application may be further configured to: after the step of analyzing the user's comprehensive recommendation index and the comprehensive score of the live broadcast content based on the machine self-learning algorithm by the preset personalized recommendation generation model to generate the user's initial personalized content list, the following steps are included:

[0024] Obtain all recommended live content items in the initial personalized content list;

[0025] Divide the recommended live content items into primary candidate content items and secondary candidate content items according to a preset correlation coefficient, and construct a primary candidate content item dataset and a secondary candidate item dataset;

[0026] The personalized recommendation generation model has a preset diversity constraint calculation formula, which is:

[0027]

[0028] in, is the preset diversity weight value, is the user-content interaction prediction value corresponding to the main candidate content item dataset, is the user-content interaction prediction value corresponding to the secondary candidate content item dataset, is the live content feature vector parameter corresponding to the main candidate content item dataset, is the live content feature vector parameter corresponding to the secondary candidate content item dataset, Used to calculate the similarity between feature vectors of live content;

[0029] When generating the initial personalized content list, live content that meets the requirements is screened based on the diversity constraints preset in the personalized recommendation generation model to be pushed to the user.

[0030] By adopting the above technical solution, by calculating the similarity between the feature vectors of the live content and adjusting the diversity weight θ, the duplication and monotony of the recommended content can be effectively avoided, and the diversity of the recommendation list can be enhanced. When faced with diverse recommended content, users are more likely to discover new points of interest and content, which helps to improve user exploration and satisfaction. This solution encourages users to be exposed to different types of live content through diversity constraints, thereby increasing users' exploration behavior and opportunities to discover new interests, thereby guiding users to be exposed to new content. By providing diverse recommended content, this solution helps prevent users from being lost due to monotonous content, thereby improving long-term user retention. Diversified content can continuously attract users' attention and increase user stickiness.

[0031] In a preferred example, the present application may be further configured to: when generating the initial personalized content list, after the step of screening the live content that meets the diversity constraint conditions preset in the personalized recommendation generation model and pushing it to the user, the following steps are included:

[0032] The preset user expectation model analyzes the diversity constraint conditions based on a machine self-learning algorithm to generate user predicted response data;

[0033] The user predicted reaction data includes predicted behavioral reaction data and predicted emotional behavior data, and the user predicted reaction data is used to determine the relevance between the pushed live broadcast content and the user's actual expected content.

[0034] By adopting the above technical solutions, user predicted reaction data can evaluate the potential effects of pushed content in advance, help the recommendation system optimize push strategies, improve the relevance of pushed content and the match between users' actual expectations, thereby improving user satisfaction and user stickiness of the platform.

[0035] In a preferred example, the present application may be further configured as follows: before the step of analyzing the diversity constraint condition based on the machine self-learning algorithm in the preset user expectation model to generate user predicted response data, the following steps are included:

[0036] Acquire historical behavior data of the user, and associate the historical behavior data with corresponding live content data to construct a historical behavior data set;

[0037] The preset user behavior analysis model analyzes the historical behavior data set based on a machine self-learning algorithm to generate a user behavior tendency value, wherein the user behavior tendency value is used to represent the stickiness value between the user and the live broadcast content;

[0038] The preset video quality analysis model analyzes the live content in the initial personalized content list based on a machine self-learning algorithm to generate a content quality index;

[0039] Obtain behavioral interaction features from historical behavior datasets, and extract and classify the interactive behavior features of users watching similar videos to analyze the time difference in users' interaction when watching similar content.

[0040] In a preferred example, the present application may be further configured as follows: a preset user interaction analysis model analyzes the user interaction dataset based on a machine self-learning algorithm to generate recommended adjustment parameters, wherein the recommended adjustment parameters are used to modify the comprehensive recommendation index to generate a modified comprehensive recommendation index. The step includes the following steps:

[0041] Acquire actual user interaction data when the user watches the live broadcast, wherein the actual user interaction data includes actual behavioral interaction data and actual emotional interaction data;

[0042] Comparing the actual user interaction data with the predicted user response data to generate user behavior deviation data;

[0043] The user interaction analysis model analyzes the user behavior deviation data to generate user intention content data, and generates recommended adjustment parameters based on the user intention content data.

[0044] In a preferred example, the present application may be further configured as follows: in the step of obtaining real-time live content data, a preset content classification model performs a comprehensive content scoring on the live content data based on a machine self-learning algorithm, including the following steps:

[0045] Obtain all live content data to construct a live content dataset;

[0046] Extracting spatiotemporal feature vectors of video clips in live content data through 3D convolutional neural networks;

[0047] Generate acoustic feature vectors in live content data through Mel-spectrogram conversion and bidirectional GRU network;

[0048] Based on the pre-trained language model, semantic encoding is performed on the barrage text in the live content data to obtain a text feature vector;

[0049] The preset dynamic gated fusion model performs weighted aggregation on the spatiotemporal feature vector, the acoustic feature vector, and the text feature vector to generate a comprehensive content feature vector;

[0050] The content classification model analyzes the comprehensive content feature vector based on a machine self-learning algorithm to generate a content quality score, wherein the content comprehensive score is used to match the comprehensive recommendation index corresponding to the user to generate recommended live content.

[0051] Secondly, the above-mentioned invention objectives of this application are achieved through the following technical solutions:

[0052] An artificial intelligence-based Internet live broadcast content recommendation device, the device comprising: a user parameter acquisition unit for acquiring user behavior parameters, live broadcast content feature parameters, and context environment parameters in real time;

[0053] A comprehensive recommendation index generating unit is configured to pre-set a content dynamic weighted calculation model to perform a fusion calculation on the user behavior parameters, live content feature parameters, and context environment parameters, and generate a comprehensive recommendation index;

[0054] A content comprehensive scoring unit is used to obtain real-time live content data, and a preset content classification model performs a content comprehensive scoring on the live content data based on a machine self-learning algorithm;

[0055] An initial personalized content list generating unit, configured to be pre-configured with a personalized recommendation generating model to analyze the user's comprehensive recommendation index and the comprehensive score of the live content based on a machine self-learning algorithm to generate an initial personalized content list for the user;

[0056] A user interaction data set construction unit is used to obtain user interaction data when a user watches a live broadcast, and associate the user interaction data with the watched live broadcast content based on a preset common timeline to construct a user interaction data set;

[0057] a recommendation adjustment parameter generating unit, configured to be pre-configured with a user interaction analysis model to analyze the user interaction dataset based on a machine self-learning algorithm to generate recommendation adjustment parameters, wherein the recommendation adjustment parameters are used to modify the comprehensive recommendation index to generate a modified comprehensive recommendation index;

[0058] The modified personalized content list generating unit is configured to be pre-configured with a content adjustment model to adjust the initial personalized content list based on the recommendation adjustment parameter to generate a modified personalized content list, and recommend live content to the user based on the modified personalized content list.

[0059] Thirdly, the above-mentioned purpose of this application is achieved through the following technical solutions:

[0060] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned method for recommending Internet live broadcast content based on artificial intelligence are implemented.

[0061] Fourthly, the above-mentioned purpose of the present application is achieved through the following technical solutions:

[0062] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for recommending Internet live broadcast content based on artificial intelligence.

[0063] In summary, this application includes at least one of the following beneficial technical effects:

[0064] 1. This invention acquires user behavior parameters, live content feature parameters, and context parameters in real time, and combines them with a dynamic weighted calculation model to generate a comprehensive recommendation index. This achieves deep integration of multi-dimensional data, more comprehensively reflecting user needs and the actual situation of live content, thereby generating more accurate recommendation results. Most existing recommendation systems lack dynamic adjustment mechanisms and are unable to optimize based on real-time user feedback. This invention, by introducing user interaction data sets and recommendation adjustment parameters, can correct the initial recommendation index in real time, further optimize recommended content, and make it more in line with users' real-time preferences.

[0065] 2. The calculation formula of the comprehensive recommendation index realizes the dynamic weighted calculation of multi-dimensional data by integrating user behavior parameters, live content feature parameters and context environment parameters, thereby significantly improving the relevance and accuracy of the recommendation system. This comprehensive consideration can not only more accurately capture the user's current preferences and needs and improve the relevance of recommended content, but also significantly enhance the user experience, making users feel more intimate and satisfactory service. The dynamic weighted calculation model can be adjusted according to the user's real-time behavior and preference changes, so that the recommended content is more in line with the user's personalized needs, while improving the platform's operational efficiency, and increasing the exposure and viewing rate of live content through accurate recommendations. In addition, the introduction of information entropy parameters and dynamic correlation coefficients promotes the diversification of content, avoids the homogenization of recommended content, and helps users discover more interesting and valuable content, time decay. The consideration of factors and contextual environment parameters enhances the adaptability of the recommendation system, enabling it to maintain good performance in different network conditions, time periods and user groups. The automated recommendation algorithm reduces manual intervention and operating costs, improves user satisfaction, reduces user churn, and thus reduces the marketing cost of acquiring new users. Real-time monitoring and analysis of user behavior data enhances the security of the platform. At the same time, the dynamic weighted calculation model can be adjusted according to the latest laws, regulations and policy requirements to ensure the compliance of recommended content. Accurate recommendation algorithms provide the platform with more business opportunities, such as advertising and content cooperation, promote the platform's technological innovation and development, and improve the platform's competitiveness. In summary, through its innovative design, the formula of this solution has brought comprehensive performance improvements to the recommendation system of the live broadcast industry and provided users with a richer and more personalized viewing experience.

[0066] 3. By calculating the similarity between the feature vectors of live content and adjusting the diversity weight θ, we can effectively avoid duplication and monotony in recommended content, thereby enhancing the diversity of the recommendation list. When faced with diverse recommended content, users are more likely to discover new points of interest and content, which helps improve user exploration and satisfaction. This solution uses diversity constraints to encourage users to access different types of live content, thereby increasing their exploration behavior and opportunities to discover new interests, thereby guiding users to access new content. By providing diverse recommended content, this solution helps prevent user churn due to monotonous content, thereby improving long-term user retention. Diverse content can continuously attract users' attention and increase user stickiness.

[0067] 4. User predicted response data can evaluate the potential effects of pushed content in advance, help the recommendation system optimize push strategies, improve the relevance of pushed content and the match between users' actual expectations, thereby improving user satisfaction and platform user stickiness. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a flow chart of an artificial intelligence-based Internet live broadcast content recommendation method in one embodiment of the present application;

[0069] Figure 2 This is a principle block diagram of an Internet live broadcast content recommendation device based on artificial intelligence in one embodiment of the present application;

[0070] Figure 3 It is a schematic diagram of an electronic device in an embodiment of the present application.

[0071] Figure Number:

[0072] 1. User parameter acquisition unit; 2. Comprehensive recommendation index generation unit; 3. Content comprehensive scoring unit; 4. Initial personalized content list generation unit; 5. User interaction data set construction unit; 6. Recommendation adjustment parameter generation unit; 7. Modified personalized content list generation unit. DETAILED DESCRIPTION

[0073] The present application is further described in detail below with reference to the accompanying drawings.

[0074] In one embodiment, if Figure 1 As shown, this application discloses an Internet live broadcast content recommendation method based on artificial intelligence, which specifically includes the following steps:

[0075] S10: Acquire user behavior parameters, live content feature parameters, and context parameters in real time;

[0076] Among them, user behavior parameters include user stay time conversion rate, cross-live room migration frequency and real-time interaction heat value;

[0077] The live content feature parameters include the host’s voiceprint feature vector, picture color dynamic entropy, and topic tag density;

[0078] The context parameters include the network delay compensation coefficient and the period traffic attenuation factor;

[0079] Specifically, in the embodiment of the present application, the user behavior parameters, live content feature parameters, and context environment parameters are exemplified as follows:

[0080] User behavior parameters: User A stays in the current live broadcast room for 30 minutes, during which the frequency of migration across live broadcast rooms is 0 (no switching between live broadcast rooms), and the real-time interaction heat value (such as the frequency of likes, comments, shares, etc.) is 5 times per minute.

[0081] Live content characteristic parameters: The host's voiceprint feature vector is extracted using voiceprint recognition technology, the picture color dynamic entropy is 0.8 (indicating that the picture is colorful and dynamically changing), and the topic tag density is 5 related topic tags appearing per minute.

[0082] Context parameters: The current network delay compensation coefficient is 1.2 (indicating that the network delay is high and compensation is required), and the period traffic attenuation factor is 0.8 (indicating that the current traffic is at its peak and the live broadcast traffic has decreased).

[0083] It should be noted that by obtaining multi-dimensional parameters in real time, we can comprehensively capture users' behavioral preferences, the characteristics of live broadcast content, and the influence of the external environment, providing a rich data foundation for subsequent accurate recommendations.

[0084] S20: A preset content dynamic weighted calculation model performs a fusion calculation on the user behavior parameters, live content feature parameters, and context environment parameters, and generates a comprehensive recommendation index;

[0085] Specifically, the content dynamic weighted calculation model performs weighted calculations on the above parameters based on preset weight assignments. For example, the user behavior parameter weight is 0.4, the live content feature parameter weight is 0.3, and the context environment parameter weight is 0.3. The result is a comprehensive recommendation index for user A of 80 points (out of 100 points). In this embodiment of the application, the sum of the weights of the user behavior parameter weight, the live content feature parameter weight, and the context environment parameter weight is 1.

[0086] S30: Real-time live content data is acquired, and a preset content classification model is used to perform a comprehensive content scoring on the live content data based on a machine self-learning algorithm;

[0087] Specifically, the system captures real-time data from the current livestream, including the host's explanation, product display images, and interactive content. The content classification model analyzes this data using a machine learning algorithm and determines that the livestream's overall content score is 85 out of 100.

[0088] S40: The preset personalized recommendation generation model analyzes the user's comprehensive recommendation index and the comprehensive score of the live broadcast content based on a machine self-learning algorithm to generate an initial personalized content list for the user;

[0089] Specifically, the personalized recommendation generation model analyzes user A's comprehensive recommendation index of 80 points and the comprehensive score of the live broadcast content of 85 points, and combines the user's historical viewing history and preferences to generate an initial personalized content list. For example, the initial personalized content list includes 3 e-commerce live broadcasts, 2 entertainment live broadcasts, and 1 knowledge popularization live broadcast.

[0090] S50: Obtaining user interaction data when the user watches the live broadcast, and associating the user interaction data with the watched live broadcast content based on a preset common timeline to construct a user interaction dataset;

[0091] Specifically, user A performs multiple likes, comments, and purchases while watching a livestream. The system associates this interaction data with the timeline of the livestream content to construct a user interaction dataset. For example, in the 10th minute of the livestream, user A likes a product displayed by the host; in the 20th minute, user A comments on the product's price-performance ratio. The system records user A's behavior data to construct a user interaction dataset.

[0092] S60: Analyzing the user interaction data set by a preset user interaction analysis model based on a machine self-learning algorithm to generate a recommended adjustment parameter, wherein the recommended adjustment parameter is used to modify the comprehensive recommendation index to generate a modified comprehensive recommendation index;

[0093] For example, the user interaction analysis model analyzed User A's interaction dataset and found that the user interacted more frequently with product displays and less frequently with advertisements. The model generated recommendation adjustment parameters, increasing the weight of live content related to product displays and decreasing the weight of advertising-related content. The resulting overall recommendation index was 82 points.

[0094] S70: The preset content adjustment model adjusts the initial personalized content list based on the recommendation adjustment parameter to generate a revised personalized content list, and recommends live content to the user based on the revised personalized content list;

[0095] For example, the content adjustment model adjusts the initial personalized content list based on the recommendation adjustment parameters, increasing the proportion of live broadcasts related to product demonstrations and reducing advertising-related content. This ultimately generates a revised personalized content list, recommending four product demonstration live broadcasts, one entertainment live broadcast, and one knowledge and science live broadcast to user A.

[0096] For steps S10-S70, through the real-time acquisition, fusion calculation and dynamic adjustment mechanism of multi-dimensional data, accurate and personalized recommendations for live content are achieved, which significantly improves user experience and platform operation efficiency. Existing live content recommendation systems mostly rely on single-dimensional data, such as only based on users' historical viewing records or simple interaction data, and lack comprehensive consideration of live content characteristics and contextual environment, resulting in inaccurate recommendation results and difficulty in meeting users' real-time needs. The present invention obtains user behavior parameters, live content feature parameters and contextual environment parameters in real time, and combines a dynamic weighted calculation model to generate a comprehensive recommendation index, thereby achieving deep fusion of multi-dimensional data, which can more comprehensively reflect user needs and the actual situation of live content, thereby generating more accurate recommendation results.

[0097] In addition, most existing recommendation systems lack a dynamic adjustment mechanism and are unable to be optimized based on real-time user feedback. The present invention, by introducing user interaction data sets and recommendation adjustment parameters, can correct the initial recommendation index in real time and further optimize the recommended content to make it more in line with the user's real-time preferences. This dynamic adjustment mechanism not only improves the accuracy and personalization of recommendations, but also enhances the interactivity between users and the platform, thereby increasing user engagement.

[0098] This invention also deeply expands the application of artificial intelligence technology. Existing technologies have limited the application of artificial intelligence to live content recommendations. However, this invention achieves a deep application of artificial intelligence technology by using machine learning algorithms to classify, score, and provide personalized recommendations for live content. This deep application not only enhances the intelligence level of the recommendation system, but also improves user experience and platform operational efficiency, while reducing the need for manual intervention and operating costs.

[0099] To sum up, the present invention has made significant improvements to the existing technology through multi-dimensional data fusion, dynamic adjustment mechanism and in-depth application of artificial intelligence technology, effectively solving the limitations of the existing technology, with outstanding substantive characteristics and significant progress, and can provide Internet live broadcast platforms with more efficient, accurate and personalized recommendation services, thereby promoting the technological development of the live broadcast industry.

[0100] In step S20: a preset content dynamic weighted calculation model performs a fusion calculation on the user behavior parameters, live content feature parameters, and context environment parameters to generate a comprehensive recommendation index. The content dynamic weighted calculation model presets a comprehensive recommendation index calculation formula as follows:

[0101] ,

[0102] + + =1;

[0103] in, The user behavior parameter weight is used to measure the importance of the user's own preferences. The weight of content feature parameters is used to measure the criticality of content quality. The context parameter weight is used to measure the adaptability of the real-time scene. is the information entropy parameter, is the normalized user parameter, is the dynamic correlation coefficient, is the context parameter, is the time decay factor, t is the time parameter, i is the i-th primary candidate content item, j is the j-th secondary candidate content item, To standardize content parameters.

[0104] The formula in this solution provides a dynamic weighted calculation model that comprehensively considers user behavior, live content characteristics, and context. Its beneficial effects are mainly reflected in the following aspects:

[0105] Multi-dimensional parameter fusion: The formula comprehensively captures all factors influencing user preferences by integrating user behavior parameters, live content feature parameters, and contextual environment parameters. This multi-dimensional parameter fusion helps generate a more accurate comprehensive recommendation index, thereby improving the accuracy of the recommendation system and user satisfaction. It should be noted that existing technologies lack the rational application of live content feature parameters and contextual environment parameters. The multi-dimensional fusion of this solution makes the recommended content more in line with user needs and guides users, thereby improving the user experience.

[0106] Dynamic weight adjustment: By introducing weight parameters α, β, and γ, the formula allows for dynamic adjustment of the importance of different parameters. This dynamic weight adjustment mechanism enables the recommendation system to flexibly adjust recommendation strategies based on different user behaviors and preferences, as well as changes in live content and environment, enhancing the personalization and adaptability of recommendations.

[0107] Application of time decay factor: Time decay factor in formula This is used to weight contextual parameters, taking into account the impact of environmental factors that change over time. This design helps the recommendation system respond promptly to environmental changes such as network conditions and user activity, thereby improving the real-time and effectiveness of recommendations.

[0108] Normalization: User parameters are normalized to eliminate the impact of different parameter dimensions, making weighted calculations fairer and more accurate. This normalization helps improve the stability and robustness of the model, making recommendation results more reliable.

[0109] Introduction of information entropy parameter: Information entropy parameter The introduction of helps to measure the uncertainty and diversity of user behavior. By considering the randomness and complexity of user behavior, the formula can better capture the user's potential needs and preferences, thereby improving the diversity and novelty of recommendations.

[0110] Application of the dynamic correlation coefficient: The dynamic correlation coefficient μj is used to measure the correlation between live content features. This design helps the recommendation system discover potential connections between content, thereby generating richer and more coherent recommendation lists, improving user exploration interest and satisfaction.

[0111] In summary, the formula in this scheme can generate a more accurate, personalized and adaptable comprehensive recommendation index by comprehensively considering user behavior, live content characteristics and contextual environment, and introducing dynamic weight adjustment, time decay factor, normalization processing and other mechanisms. The comprehensive recommendation index calculation formula realizes dynamic weighted calculation of multi-dimensional data by integrating user behavior parameters, live content characteristic parameters and contextual environment parameters, thereby significantly improving the relevance and accuracy of the recommendation system. This comprehensive consideration can not only more accurately capture the user's current preferences and needs and improve the relevance of recommended content, but also significantly enhance the user experience and make users feel more intimate and satisfactory service. The dynamic weighted calculation model can be adjusted according to the user's real-time behavior and preference changes, so that the recommended content is more in line with the user's personalized needs, while improving the platform's operational efficiency and increasing the exposure and viewing rate of live content through accurate recommendations. In addition, the introduction of information entropy parameters and dynamic correlation coefficients promotes It promotes content diversification, avoids the homogeneity of recommended content, and helps users discover more interesting and valuable content. The consideration of time decay factor and contextual environment parameters enhances the adaptability of the recommendation system, enabling it to maintain good performance in different network conditions, time periods and user groups. The automated recommendation algorithm reduces manual intervention and operating costs, improves user satisfaction, reduces user churn, and thus reduces the marketing cost of acquiring new users. Real-time monitoring and analysis of user behavior data enhances the security of the platform. At the same time, the dynamic weighted calculation model can be adjusted according to the latest laws, regulations and policy requirements to ensure the compliance of recommended content. Accurate recommendation algorithms provide the platform with more business opportunities, such as advertising, content cooperation, etc., promote the platform's technological innovation and development, and improve the platform's competitiveness. In summary, through its innovative design, the formula of this solution has brought comprehensive performance improvements to the recommendation system of the live broadcast industry and provided users with a richer and more personalized viewing experience.

[0112] After the step S40 of analyzing the user's comprehensive recommendation index and the comprehensive score of the live broadcast content based on a machine self-learning algorithm to generate an initial personalized content list for the user, the following steps are included:

[0113] S41: Obtain all recommended live content items in the initial personalized content list;

[0114] S42: Dividing the recommended live content items into primary candidate content items and secondary candidate content items according to a preset correlation coefficient, and constructing a primary candidate content item dataset and a secondary candidate item dataset;

[0115] The personalized recommendation generation model has a preset diversity constraint calculation formula, which is:

[0116]

[0117] in, is the preset diversity weight value, is the user-content interaction prediction value corresponding to the main candidate content item dataset, is the user-content interaction prediction value corresponding to the secondary candidate content item dataset, is the live content feature vector parameter corresponding to the main candidate content item dataset, is the live content feature vector parameter corresponding to the secondary candidate content item dataset, Used to calculate the similarity between feature vectors of live content;

[0118] S43: When generating an initial personalized content list, live broadcast content that meets the requirements is selected based on the diversity constraint conditions preset in the personalized recommendation generation model to be pushed to the user.

[0119] Existing technologies often overemphasize users' past behavior or preferences, leading to homogeneous recommended content and a lack of novelty. This solution calculates the similarity between live content feature vectors and, under the adjustment of the diversity weight θ, effectively avoids duplication and monotony in recommended content, thereby enhancing the diversity of the recommendation list. When faced with diverse recommended content, users are more likely to discover new points of interest and content, which helps improve user exploration and satisfaction. This solution, through diversity constraints, encourages users to engage with different types of live content, thereby increasing their exploration behavior and opportunities to discover new interests, thereby guiding users to engage with new content. By providing diverse recommended content, this solution helps prevent user churn due to monotonous content, thereby improving long-term user retention. Diverse content can continuously attract users' attention and increase user stickiness. Furthermore, through diversity constraints, this solution can more effectively distribute different types of live content to appropriate user groups, thereby improving the efficiency and effectiveness of content distribution. This helps content creators reach a wider audience and allows users to discover more high-quality content. By encouraging diverse content recommendations, this solution helps promote the healthy development of the live content ecosystem. Diverse recommendation mechanisms can provide exposure opportunities for different types of content creators, thereby enriching the entire content ecosystem.

[0120] In S43: when generating the initial personalized content list, after the step of screening the live content that meets the requirements based on the diversity constraint conditions preset in the personalized recommendation generation model to push to the user, the following steps are included:

[0121] S431: Acquire historical behavior data of the user, and associate the historical behavior data with corresponding live content data to construct a historical behavior data set;

[0122] Specifically, suppose user B watched five live game broadcasts, three live music broadcasts, and two educational lectures in the past week. The system associates this behavior data with the corresponding live content tags (such as game, music, and education) to construct a historical behavior dataset.

[0123] S432: Analyzing the historical behavior data set using a pre-set user behavior analysis model based on a machine self-learning algorithm to generate a user behavior tendency value, where the user behavior tendency value is used to represent the stickiness value between the user and the live broadcast content;

[0124] Specifically, the user behavior analysis model analyzes user B's historical behavior dataset and finds that the user watches game live broadcasts for the longest time and interacts most frequently, thus generating a high game live broadcast propensity value (such as 0.7), while the propensity values ​​for music live broadcasts and educational lectures are relatively low (such as 0.2 and 0.1).

[0125] User behavior tendency values ​​can quantify users' preferences for different types of live content, helping the recommendation system to more accurately capture users' personalized needs and improve the accuracy of recommendations and user satisfaction.

[0126] S433: Analyzing the live content in the initial personalized content list using a preset video quality analysis model based on a machine self-learning algorithm to generate a content quality index;

[0127] Specifically, the video quality analysis model analyzes the live broadcast content in the initial personalized content list, taking into account factors such as the host's professionalism, picture clarity, and interactive enthusiasm, and generates a content quality index for each live broadcast. For example, the content quality index of game live broadcast A is 85 points, and the content quality index of music live broadcast B is 75 points.

[0128] S434: Obtaining behavioral interaction features from the historical behavior dataset, and extracting and classifying the interactive behavior features of users watching similar videos to analyze the time difference in users' interaction when watching similar content;

[0129] Specifically, the system extracts the behavioral characteristics of user B watching live game broadcasts from the historical behavior dataset and finds that there is usually a time interval of 2-3 days between users watching two live game broadcasts. The analysis of this interaction time difference helps to understand the user's viewing habits and rhythm. It should be noted that the unit of the time interval can be days, minutes, hours, or other time quantification units.

[0130] By analyzing the interaction time difference between users who watch similar content, the recommendation system can better grasp the timing of push notifications, avoid user fatigue caused by excessive push notifications, and improve the acceptance and effectiveness of push notifications.

[0131] S435: Analyzing the diversity constraint condition using a preset user expectation model based on a machine self-learning algorithm to generate user predicted response data;

[0132] The user predicted reaction data includes predicted behavioral reaction data and predicted emotional behavior data, and the user predicted reaction data is used to determine the relevance between the pushed live broadcast content and the user's actual expected content;

[0133] Specifically, by analyzing the interaction time difference between users watching similar content, the recommendation system can better grasp the timing of push notifications, avoid excessive push notifications that lead to user fatigue, and improve the acceptance and effectiveness of push notifications.

[0134] In summary, the user predicted reaction data in steps S431-S435 can evaluate the potential effect of the pushed content in advance, help the recommendation system optimize the push strategy, improve the relevance of the pushed content and the match between the user's actual expectations, thereby improving user satisfaction and user stickiness of the platform.

[0135] In step S30: acquiring real-time live content data, a preset content classification model performs a comprehensive content scoring on the live content data based on a machine self-learning algorithm, including the following steps:

[0136] S301: Acquire all live content data to construct a live content data set;

[0137] S302: extracting spatiotemporal feature vectors of video clips in the live content data through a 3D convolutional neural network;

[0138] S303: Generate acoustic feature vectors in the live content data through Mel-spectrogram conversion and bidirectional GRU network;

[0139] S304: semantically encode the barrage text in the live content data based on the pre-trained language model to obtain a text feature vector;

[0140] S305: A preset dynamic gated fusion model performs weighted aggregation on the spatiotemporal feature vector, the acoustic feature vector, and the text feature vector to generate a comprehensive content feature vector;

[0141] The content classification model analyzes the comprehensive content feature vector based on a machine self-learning algorithm to generate a content quality score, wherein the content comprehensive score is used to match the comprehensive recommendation index corresponding to the user to generate recommended live content.

[0142] It should be noted that, in the embodiment of the present application, the spatiotemporal feature vector is , the acoustic eigenvector is , the text feature vector is , the calculation formula of the comprehensive content feature vector is

[0143] in is the sigmoid function, 、 、 is a learnable weight matrix, and ∥ represents a vector concatenation operation.

[0144] The corresponding content quality score calculation formula is: ,in is the internal soft category adjustment factor corresponding to the preset i-th category content (for example, educational content =1.5 Entertainment content ), U is the weight matrix of the fully connected layer, and ReLU is the rectified linear unit function.

[0145] It should be noted that in other embodiments, the hotspot parameter values ​​corresponding to the live broadcast can be increased, for example, based on the number and frequency of sharing of the live broadcast in the pre-interval as a reference, so as to improve the accuracy of the live broadcast content rating, push high-quality live broadcast content to users, and improve the user experience.

[0146] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0147] In one embodiment, an artificial intelligence-based Internet live content recommendation device is provided, which corresponds to the artificial intelligence-based Internet live content recommendation method in the above embodiment. Figure 2 As shown, the Internet live broadcast content recommendation device based on artificial intelligence includes a user parameter acquisition unit 1, which is used to obtain user behavior parameters, live broadcast content feature parameters and context environment parameters in real time;

[0148] A comprehensive recommendation index generating unit 2 is configured to pre-set a content dynamic weighted calculation model to perform a fusion calculation on the user behavior parameters, live content feature parameters, and context environment parameters, and generate a comprehensive recommendation index;

[0149] A content comprehensive scoring unit 3 is used to obtain real-time live content data and perform a content comprehensive scoring on the live content data using a pre-set content classification model based on a machine self-learning algorithm;

[0150] An initial personalized content list generating unit 4 is configured to be pre-configured with a personalized recommendation generating model to analyze the user's comprehensive recommendation index and the comprehensive score of the live content based on a machine self-learning algorithm to generate an initial personalized content list for the user;

[0151] A user interaction data set construction unit 5 is configured to obtain user interaction data when a user watches a live broadcast, and associate the user interaction data with the watched live broadcast content based on a preset common timeline to construct a user interaction data set;

[0152] a recommended adjustment parameter generating unit 6, configured to be pre-configured with a user interaction analysis model to analyze the user interaction dataset based on a machine self-learning algorithm to generate recommended adjustment parameters, wherein the recommended adjustment parameters are used to modify the comprehensive recommendation index to generate a modified comprehensive recommendation index;

[0153] The modified personalized content list generating unit 7 is configured to be pre-configured with a content adjustment model to adjust the initial personalized content list based on the recommendation adjustment parameter to generate a modified personalized content list, and recommend live content to the user based on the modified personalized content list.

[0154] For the specific definition of the Internet live content recommendation device based on artificial intelligence, please refer to the definition of the Internet live content recommendation method based on artificial intelligence above, which will not be repeated here. The various modules in the above-mentioned Internet live content recommendation device based on artificial intelligence can be implemented in whole or in part by software, hardware, and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the electronic device in the form of hardware, or can be stored in the memory of the electronic device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0155] In one embodiment, an electronic device is provided. The electronic device may be a server, and its internal structure diagram may be as follows: Figure 3 As shown. The electronic device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the electronic device is used to store a database. The network interface of the electronic device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, an artificial intelligence-based Internet live content recommendation method is implemented.

[0156] In one embodiment, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following steps are performed:

[0157] Real-time acquisition of user behavior parameters, live content feature parameters, and context parameters, including:

[0158] User behavior parameters include user stay time conversion rate, cross-live room migration frequency, and real-time interaction heat value;

[0159] The live content feature parameters include the host’s voiceprint feature vector, picture color dynamic entropy, and topic tag density;

[0160] The context parameters include the network delay compensation coefficient and the period traffic attenuation factor;

[0161] The preset content dynamic weighted calculation model performs a fusion calculation on the user behavior parameters, live content feature parameters and context environment parameters, and generates a comprehensive recommendation index;

[0162] Acquire real-time live content data, and use a pre-set content classification model to perform a comprehensive content scoring on the live content data based on a machine self-learning algorithm;

[0163] The preset personalized recommendation generation model analyzes the user's comprehensive recommendation index and the comprehensive score of the live broadcast content based on a machine self-learning algorithm to generate an initial personalized content list for the user;

[0164] Acquire user interaction data when the user watches the live broadcast, and associate the user interaction data with the watched live broadcast content based on a preset common timeline to construct a user interaction dataset;

[0165] A preset user interaction analysis model analyzes the user interaction dataset based on a machine self-learning algorithm to generate recommended adjustment parameters, wherein the recommended adjustment parameters are used to modify the comprehensive recommendation index to generate a modified comprehensive recommendation index;

[0166] The preset content adjustment model adjusts the initial personalized content list based on the recommendation adjustment parameter to generate a revised personalized content list, and recommends live content to the user based on the revised personalized content list.

[0167] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0168] Real-time acquisition of user behavior parameters, live content feature parameters, and context parameters, including:

[0169] User behavior parameters include user stay time conversion rate, cross-live room migration frequency, and real-time interaction heat value;

[0170] The live content feature parameters include the host’s voiceprint feature vector, picture color dynamic entropy, and topic tag density;

[0171] The context parameters include the network delay compensation coefficient and the period traffic attenuation factor;

[0172] The preset content dynamic weighted calculation model performs a fusion calculation on the user behavior parameters, live content feature parameters and context environment parameters, and generates a comprehensive recommendation index;

[0173] Acquire real-time live content data, and use a pre-set content classification model to perform a comprehensive content scoring on the live content data based on a machine self-learning algorithm;

[0174] The preset personalized recommendation generation model analyzes the user's comprehensive recommendation index and the comprehensive score of the live broadcast content based on a machine self-learning algorithm to generate an initial personalized content list for the user;

[0175] Acquire user interaction data when the user watches the live broadcast, and associate the user interaction data with the watched live broadcast content based on a preset common timeline to construct a user interaction dataset;

[0176] A preset user interaction analysis model analyzes the user interaction dataset based on a machine self-learning algorithm to generate recommended adjustment parameters, wherein the recommended adjustment parameters are used to modify the comprehensive recommendation index to generate a modified comprehensive recommendation index;

[0177] The preset content adjustment model adjusts the initial personalized content list based on the recommendation adjustment parameter to generate a revised personalized content list, and recommends live content to the user based on the revised personalized content list.

[0178] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0179] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0180] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. An artificial intelligence-based method for recommending live internet content, characterized in that: The method comprises the steps of: Real-time acquisition of user behavior parameters, live content feature parameters, and context parameters, including: User behavior parameters include user stay time conversion rate, cross-live room migration frequency, and real-time interaction heat value; The live content feature parameters include the host’s voiceprint feature vector, picture color dynamic entropy, and topic tag density; The context parameters include the network delay compensation coefficient and the period traffic attenuation factor; The preset content dynamic weighted calculation model performs a fusion calculation on the user behavior parameters, live content feature parameters and context environment parameters, and generates a comprehensive recommendation index; Acquire real-time live content data, and use a pre-set content classification model to perform a comprehensive content scoring on the live content data based on a machine self-learning algorithm; The preset personalized recommendation generation model analyzes the user's comprehensive recommendation index and the comprehensive score of the live broadcast content based on a machine self-learning algorithm to generate an initial personalized content list for the user; Acquire user interaction data when the user watches the live broadcast, and associate the user interaction data with the watched live broadcast content based on a preset common timeline to construct a user interaction dataset; A preset user interaction analysis model analyzes the user interaction dataset based on a machine self-learning algorithm to generate recommended adjustment parameters, wherein the recommended adjustment parameters are used to modify the comprehensive recommendation index to generate a modified comprehensive recommendation index; The preset content adjustment model adjusts the initial personalized content list based on the recommendation adjustment parameter to generate a revised personalized content list, and recommends live content to the user based on the revised personalized content list.

2. The method for recommending Internet live broadcast content based on artificial intelligence according to claim 1, characterized in that: In the step of integrating and calculating the user behavior parameters, live content feature parameters, and context environment parameters using a preset content dynamic weighted calculation model to generate a comprehensive recommendation index, the content dynamic weighted calculation model presets a comprehensive recommendation index calculation formula as follows: ， + + =1； in, The user behavior parameter weight is used to measure the importance of the user's own preferences. The weight of content feature parameters is used to measure the criticality of content quality. The context parameter weight is used to measure the adaptability of the real-time scene. is the information entropy parameter, is the normalized user parameter, is the dynamic correlation coefficient, is the context parameter, is the time decay factor, t is the time parameter, i is the i-th primary candidate content item, j is the j-th secondary candidate content item, To standardize content parameters.

3. The method for recommending Internet live broadcast content based on artificial intelligence according to claim 2, characterized in that: After the preset personalized recommendation generation model analyzes the user's comprehensive recommendation index and the comprehensive score of the live broadcast content based on the machine self-learning algorithm to generate the user's initial personalized content list, the following steps are included: Obtain all recommended live content items in the initial personalized content list; Divide the recommended live content items into primary candidate content items and secondary candidate content items according to a preset correlation coefficient, and construct a primary candidate content item dataset and a secondary candidate item dataset; The personalized recommendation generation model has a preset diversity constraint calculation formula, which is: in, is the preset diversity weight value, is the user-content interaction prediction value corresponding to the main candidate content item dataset, is the user-content interaction prediction value corresponding to the secondary candidate content item dataset, is the live content feature vector parameter corresponding to the main candidate content item dataset, is the live content feature vector parameter corresponding to the secondary candidate content item dataset, Used to calculate the similarity between feature vectors of live content; When generating the initial personalized content list, live content that meets the requirements is screened based on the diversity constraints preset in the personalized recommendation generation model to be pushed to the user.

4. The method for recommending Internet live broadcast content based on artificial intelligence according to claim 3, characterized in that: When generating the initial personalized content list, after the step of screening the live content that meets the diversity constraint conditions preset in the personalized recommendation generation model to push it to the user, the following steps are included: The preset user expectation model analyzes the diversity constraint conditions based on a machine self-learning algorithm to generate user predicted response data; The user predicted reaction data includes predicted behavioral reaction data and predicted emotional behavior data, and the user predicted reaction data is used to determine the relevance between the pushed live broadcast content and the user's actual expected content.

5. The method for recommending Internet live broadcast content based on artificial intelligence according to claim 4, characterized in that: Before the step of analyzing the diversity constraint conditions based on the machine self-learning algorithm by the preset user expectation model to generate user predicted response data, the following steps are included: Acquire historical behavior data of the user, and associate the historical behavior data with corresponding live content data to construct a historical behavior data set; The preset user behavior analysis model analyzes the historical behavior data set based on a machine self-learning algorithm to generate a user behavior tendency value, wherein the user behavior tendency value is used to represent the stickiness value between the user and the live broadcast content; The preset video quality analysis model analyzes the live content in the initial personalized content list based on a machine self-learning algorithm to generate a content quality index; Obtain behavioral interaction features from historical behavior datasets, and extract and classify the interactive behavior features of users watching similar videos to analyze the time difference in users' interaction when watching similar content.

6. The method for recommending Internet live broadcast content based on artificial intelligence according to claim 4, characterized in that: The step of analyzing the user interaction data set based on a machine self-learning algorithm by a preset user interaction analysis model to generate recommended adjustment parameters, wherein the recommended adjustment parameters are used to correct the comprehensive recommendation index to generate a corrected comprehensive recommendation index, includes the following steps: Acquire actual user interaction data when the user watches the live broadcast, wherein the actual user interaction data includes actual behavioral interaction data and actual emotional interaction data; Comparing the actual user interaction data with the predicted user response data to generate user behavior deviation data; The user interaction analysis model analyzes the user behavior deviation data to generate user intention content data, and generates recommended adjustment parameters based on the user intention content data.

7. The method for recommending Internet live broadcast content based on artificial intelligence according to claim 1, characterized in that: In the step of obtaining real-time live content data and performing a comprehensive content scoring on the live content data using a pre-set content classification model based on a machine self-learning algorithm, the following steps are included: Obtain all live content data to construct a live content dataset; Extracting spatiotemporal feature vectors of video clips in live content data through 3D convolutional neural networks; Generate acoustic feature vectors in live content data through Mel-spectrogram conversion and bidirectional GRU network; Based on the pre-trained language model, semantic encoding is performed on the barrage text in the live content data to obtain a text feature vector; The preset dynamic gated fusion model performs weighted aggregation on the spatiotemporal feature vector, the acoustic feature vector, and the text feature vector to generate a comprehensive content feature vector; The content classification model analyzes the comprehensive content feature vector based on a machine self-learning algorithm to generate a content quality score, wherein the content comprehensive score is used to match the comprehensive recommendation index corresponding to the user to generate recommended live content.

8. An artificial intelligence-based Internet live broadcast content recommendation device, applied to the artificial intelligence-based Internet live broadcast content recommendation method according to any one of claims 1 to 7, characterized in that: The device comprises: a user parameter acquisition unit (1) for acquiring user behavior parameters, live content feature parameters and context environment parameters in real time; A comprehensive recommendation index generating unit (2) is configured to pre-set a content dynamic weighted calculation model to perform fusion calculation on the user behavior parameters, live content feature parameters and context environment parameters, and generate a comprehensive recommendation index; A content comprehensive scoring unit (3) is used to obtain real-time live content data, and a preset content classification model performs a content comprehensive scoring on the live content data based on a machine self-learning algorithm; An initial personalized content list generating unit (4) is configured to be pre-configured with a personalized recommendation generating model to analyze the user's comprehensive recommendation index and the comprehensive score of the live content based on a machine self-learning algorithm to generate the user's initial personalized content list; A user interaction data set construction unit (5) is used to obtain user interaction data when a user watches a live broadcast, and associate the user interaction data with the watched live broadcast content based on a preset common time axis to construct a user interaction data set; a recommendation adjustment parameter generating unit (6), configured to be pre-configured with a user interaction analysis model to analyze the user interaction data set based on a machine self-learning algorithm to generate recommendation adjustment parameters, wherein the recommendation adjustment parameters are used to modify the comprehensive recommendation index to generate a modified comprehensive recommendation index; A modified personalized content list generating unit (7) is configured to be pre-set with a content adjustment model to adjust the initial personalized content list based on the recommendation adjustment parameter to generate a modified personalized content list, and to recommend live content to the user based on the modified personalized content list.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method for recommending Internet live broadcast content based on artificial intelligence as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for recommending Internet live broadcast content based on artificial intelligence as described in any one of claims 1 to 7 are implemented.

Images