AIGC content recommendation method and system based on user participation degree optimization

Abstract

The invention discloses an AIGC content recommendation method and system based on user participation degree optimization, and belongs to the technical field of data intelligent analysis, and the method comprises the steps: collecting interaction data between a user and historical content, generating the participation degree of the user and the participated degree of the historical content, and screening out the interested content of the user; generating an information stream of the historical content, extracting index information in the information stream, and generating a content change curve corresponding to the interested content and an interest change curve corresponding to the user; the optimization model dynamically adjusts parameters within a preset time interval based on the content change curve and the participation degree to generate a standard recommendation model, generates an association database based on the index information, extracts optimization content from the association database and pushes the optimization content to the user; and repeatedly executing the step until the participation degree between the user and the optimized content is greater than or equal to a first preset value, and completing content recommendation. According to the invention, the reliability of content recommendation and the user participation degree can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent data analysis, and specifically relates to an AIGC content recommendation method and system based on user participation optimization. Background Art

[0002] Artificial Intelligence Generated Content (AIGC) is a process that combines multiple AI technologies to automatically generate text, images, audio, and video content. These contents can be completely original or based on existing data and patterns. Collecting user participation in AIGC-generated content and further optimizing generated content and recommendation algorithms will help achieve more accurate content recommendations.

[0003] Similar prior art includes a Chinese patent application with publication number CN106649884A, which discloses a multimedia content recommendation method based on user context analysis. According to the identity, behavior, intention and environmental big data of the user in the social media network, after analysis, the multimedia content that the target user may be interested in is recommended, so that the user can quickly and conveniently find the content he likes, and improve the accuracy of the recommendation algorithm; free the user from the massive multimedia resources, save the user's time and improve the user's experience quality in the multimedia social network. There is also a Chinese patent application with publication number CN104133820A, which discloses a content recommendation method and content recommendation device, belonging to the field of Internet technology. The method includes: analyzing the pages browsed by the user and extracting the theme keywords; obtaining a list of URLs corresponding to the theme keywords from a preset theme library; and recommending content related to the theme keywords to the user according to the content of the web page corresponding to the URL in the URL list. The invention recommends content related to themes through the content recommendation device, and can recommend content within the scope of various websites. Compared with the prior art in which the website recommends content related to themes within the scope of the website, it can expand the recommendation scope, enhance the recommendation effect and improve the recommendation efficiency. By adopting this technical solution, users can obtain the contents of related topics recommended by various websites recommended by the content recommendation device, which is very convenient to use.

[0004] However, the above-mentioned existing technologies need to obtain more information data from users, which affects the information security of users. They also need to collect the URLs browsed by users. The amount of data analysis is large, which is not conducive to the rapidity of content recommendation. In actual situations, it is necessary to generate content recommendations based on changes in user interests, so as to improve user participation and the reliability of content recommendations. Summary of the invention

[0005] To solve the above problems, the present invention provides an AIGC content recommendation method and system based on user engagement optimization to solve the problems in the prior art.

[0006] In order to achieve the above-mentioned object of the invention, the present invention proposes an AIGC content recommendation method based on user participation optimization, comprising:

[0007] S1: Collecting interaction data between the user and historical content, generating the user's engagement and the engagement of the historical content based on the interaction data, and filtering out the user's interested content from all historical content based on the engagement;

[0008] S2: Generate an information stream of the historical content based on preset tags, extract index information from the information stream based on preset description words, generate a content change curve corresponding to the interested content based on the index information, and generate an interest change curve of the user based on the content change curve;

[0009] S3: constructing an optimization model, wherein the optimization model dynamically adjusts parameters within a preset time interval based on the content change curve and the engagement level to generate a standard recommendation model, and generates an associated database based on the index information, wherein the standard recommendation model extracts optimized content from the associated database based on the interest change curve and pushes the optimized content to the user;

[0010] S4: Repeat S3 until the engagement level between the user and the optimized content is greater than or equal to a first preset value, thereby completing content recommendation.

[0011] Furthermore, the generating the information flow of the historical content based on the preset tag comprises the following steps:

[0012] The preset tags include string type, video type and image type. All the historical contents are classified into multiple subcategories based on the preset tags. If the subcategory is the string type, the summary text of the historical content is extracted. If the subcategory is the video type, the text annotation of the historical content is extracted. If the subcategory is the image type, the image tag of the historical content is extracted. The summary text, the text annotation and the image tag are respectively set as the information flow corresponding to the historical content.

[0013] Furthermore, the extracting of text annotations of the historical content comprises the following steps:

[0014] Segmenting the historical content into a plurality of sub-segments based on an audio signal, constructing a multimodal model, and acquiring multi-type data corresponding to the sub-segments based on the multimodal model, wherein the multi-type data includes video frames, background audio frames, and text information;

[0015] Based on the multimodal model, the scene information of the video frame is obtained, a content label and an emotion label are set, and it is determined whether the semantic similarity between the scene information and the text information is greater than a second preset value. If yes, the scene information and the text information are combined to generate fitting information, and the content category and the emotion category corresponding to the content label and the emotion label are respectively extracted from the fitting information; otherwise, the content category and the emotion category corresponding to the content label and the emotion label are respectively extracted from the text information and the scene information;

[0016] Establish an audio database based on preset audio tags, extract reference audio with the greatest signal waveform similarity from the audio database based on the signal waveform of the background audio frame, and set the preset audio tag corresponding to the reference audio as the audio category of the background audio frame;

[0017] The content categories, the emotion categories and the audio categories corresponding to all the sub-segments are sequentially combined and set as text annotations of the historical content.

[0018] Furthermore, the acquiring the scene information of the video frame based on the multimodal model comprises the following steps:

[0019] Collecting video samples, marking label indexes of the video samples, combining all the video samples with the corresponding label indexes to form a training sample set, and generating the multimodal model based on a neural network and feature vectors of the training sample set;

[0020] The multimodal model divides the video frame into multiple frame images based on a preset frame, calculates the image similarity between the frame image and all the video samples, generates a label index corresponding to the frame image based on the image similarity, sets it as the image information of the frame image, and sequentially combines all the image information to generate the scene information.

[0021] Furthermore, marking the label index of the video sample includes the following steps:

[0022] All the video samples are classified into labeled videos and test videos, related objects of the labeled videos are obtained based on image motion detection, keyword groups are generated based on the types of the related objects, semantic tags corresponding to the keyword groups are generated based on a preset knowledge graph, the first probability of each semantic tag in the test video is output, the semantic tag whose first probability is greater than or equal to a third preset value is extracted, set as a test tag, and the keyword group and the test tag are respectively marked as the tag index corresponding to the video sample.

[0023] Further, generating the user's interest change curve based on the content change curve comprises the following steps:

[0024] Constructing a standard content tag set, calculating the text similarity between the index information and any standard content tag in the standard content tag set, normalizing all text similarities to generate strength values ​​of the standard content tags, setting all the standard content tags as horizontal axis information, setting the strength values ​​as vertical axis values, and generating the content change curve based on the horizontal axis information and the corresponding vertical axis values;

[0025] Arrange all the contents of interest in sequence based on the timestamp of the interaction data to generate a content time series, calculate the standard deviation corresponding to the content change curve of the contents of interest, set it as the value of the content time series at the timestamp, connect all the values ​​at the timestamps in sequence, and generate the interest change curve.

[0026] Furthermore, the extracting and optimizing content includes the following steps:

[0027] Obtain content change curves corresponding to all updated contents in the associated database and set them as the first curve, cluster all first curves to generate multiple cluster categories, set the updated contents contained in each cluster category as the first content, calculate the average value of the standard deviations corresponding to all first contents, the standard recommendation model outputs a predicted value at a predicted time point based on the interest change curve and the timestamp, calculates the difference between the predicted value and the average value, sets the first content corresponding to the minimum difference as the candidate content, and sets the candidate content with the minimum standard deviation as the optimized content.

[0028] Furthermore, the change rate of each parameter in the optimization model is adjusted based on the change rate of the standard deviation of all the first contents to generate the standard recommendation model.

[0029] Furthermore, extracting the summary text of the historical content includes the following steps:

[0030] A natural language processing model is constructed to obtain text characters of the historical content based on the string type. The natural language processing model extracts semantic information from the text characters and obtains a character group with the highest word frequency in the text characters. The character group is combined with the semantic information to generate the summary text.

[0031] The present invention also provides an AIGC content recommendation system based on user participation optimization, which is used to implement the above-mentioned AIGC content recommendation method based on user participation optimization. The system mainly includes:

[0032] A content collection module collects interaction data between a user and historical content, generates the user's engagement and the engagement of the historical content based on the interaction data, and selects the user's interested content from all historical content based on the engagement;

[0033] A content analysis module, which generates an information stream of the historical content based on preset tags, extracts index information from the information stream based on preset descriptive words, generates a content change curve corresponding to the content of interest based on the index information, and generates an interest change curve of the user based on the content change curve;

[0034] A content recommendation module, constructing an optimization model, wherein the optimization model dynamically adjusts parameters within a preset time interval based on the content change curve and the engagement level to generate a standard recommendation model, and generates an associated database based on the index information, wherein the standard recommendation model extracts optimized content from the associated database based on the interest change curve and pushes it to the user;

[0035] The model optimization module repeatedly executes S3 until the engagement degree between the user and the optimized content is greater than or equal to a first preset value, thereby completing content recommendation.

[0036] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0037] The present invention first extracts the user's content of interest based on interaction data, and then analyzes the information flow of different types of historical content according to preset tags, which can accurately generate a content change curve corresponding to the content of interest, and further generate a user's interest change curve, which can accurately describe the user's browsing habits and interest change process. Finally, dynamic adjustment parameters are set to generate a standard recommendation model. The predicted value output by the standard recommendation model is used to extract optimized content in the associated database and push it to the user, which can improve the reliability of content recommendation based on the analysis of user participation in browsing records.

[0038] The present invention also uses a multimodal model to accurately analyze different types of historical content, accurately generate content change curves corresponding to the historical content, and improve the accuracy of extracting optimized content. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A flowchart of the steps of an AIGC content recommendation method based on user engagement optimization of the present invention;

[0040] Figure 2 This is a structural diagram of an AIGC content recommendation system based on user engagement optimization according to the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0042] It is understood that the terms "first", "second", etc. used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are only used to distinguish a first element from another element. For example, without departing from the scope of this application, a first xx script may be referred to as a second xx script, and similarly, a second xx script may be referred to as a first xx script.

[0043] like Figure 1 As shown, an AIGC content recommendation method based on user engagement optimization includes:

[0044] S1: Collect the interaction data between users and historical content, generate user engagement and historical content engagement based on the interaction data statistics, and filter out the user's interested content from all historical content based on the engagement.

[0045] Specifically, the user's browsing history will display the data information and browsing time of the historical content, which can indirectly obtain the user's participation. In this embodiment, the historical content refers to the content that the user has browsed, the interaction data refers to the time point and the time consumed when the user browses the historical content, and the participation refers to the ratio between the time consumed by the user to browse the historical content and the required time to browse the historical content, wherein the required time to browse the historical content can be set by calculating the average time that the historical content is browsed, and the participation refers to the ratio between the number of historical content browsed and the total number of browsed, wherein the browsing is completed means that the browsing time of the historical content is greater than or equal to the required time. For example, the required time of the historical content H1 is 5 minutes, and the user browses the historical content H1 for 1 minute, then the user's participation is 1 / 5=0.2, the total number of browsing historical content H1 is 10 times, of which the number of browsing is completed is 1 time, then the participation of the historical content H1 is 1 / 10=0.1. Interested content refers to historical content with an engagement degree greater than or equal to a first threshold among all historical content browsed by the user. For example, if the first threshold is set to 0.5, all historical content with an engagement degree greater than or equal to 0.5 are set as the user's interested content.

[0046] S2: Generate an information stream of historical content based on preset tags, extract index information from the information stream based on preset description words, generate a content change curve corresponding to the content of interest based on the index information, and generate a user interest change curve based on the content change curve.

[0047] Specifically, in this embodiment, the preset tags are tags used to classify historical content types, including but not limited to string types, video types, and image types. The information flow refers to the useful information extracted from the historical content. The index information refers to the descriptive information extracted from the information flow, including but not limited to action words, sentiment words, adjectives, and other information. The content change curve is used to describe the change law of the index information of the content of interest. The interest change curve is used to describe the change law of the user's browsing history content over time. The interest change curve can indirectly generate the change information of the user's participation.

[0048] S3: Build an optimization model. The optimization model dynamically adjusts parameters within a preset time interval based on the content change curve and engagement to generate a standard recommendation model. The associated database is generated based on the index information. The standard recommendation model extracts optimized content from the associated database based on the interest change curve and pushes it to users.

[0049] Specifically, in this embodiment, both the optimization model and the standard recommendation model are prediction models generated by the LSTM long short-term memory network, wherein the standard recommendation model is a model generated corresponding to the parameters of the optimization model after adjustment. The associated database refers to a content database generated by AIGC technology based on index information, which contains content data corresponding to a variety of index information, wherein AIGC uses machine learning or deep learning algorithms to train models and highly automatically generates a large amount of content. The optimized content is the content information that is currently pushed to the user for browsing, and the specific extraction method will be described later.

[0050] S4: Repeat S3 until the engagement degree between the user and the optimized content is greater than or equal to a first preset value, and the content recommendation is completed.

[0051] Specifically, in this embodiment, in order to improve the user's engagement in browsing optimized content, it is necessary to continuously adjust the parameters of the optimization model and improve the standard recommendation model so that the engagement between the user and the optimized content is greater than or equal to a first preset value, wherein the first preset value can be set to 0.8 to complete content recommendation based on user engagement optimization.

[0052] As a preferred technical solution of the present invention, generating an information stream of historical content based on preset tags includes the following steps:

[0053] The preset tags include string type, video type and image type. All historical contents are classified into multiple subcategories based on the preset tags. If the subcategory is a string type, the summary text of the historical content is extracted. If the subcategory is a video type, the text annotation of the historical content is extracted. If the subcategory is an image type, the image tag of the historical content is extracted. The summary text, text annotation and image tag are respectively set as the information flow corresponding to the historical content.

[0054] Specifically, in this embodiment, the string type refers to the content type in the form of a string text, the video type refers to the content type in the form of a video, the image type refers to the content type in the form of a picture, and the subcategory refers to a plurality of historical contents of the same form composed of all historical contents classified by preset labels. Summary text refers to the characters generated by summarizing and reasoning after extracting all the strings contained in the historical content, text annotation refers to the characters generated by reading the video information of the historical content and performing contextual annotation of all the video content, and image label refers to the characters generated by identifying the image information of the historical content and summarizing the image information. Through three different types of preset labels, the corresponding historical content can be accurately processed, thereby extracting the information flow of the historical content and improving the integrity and accuracy of the information flow.

[0055] As a preferred technical solution of the present invention, extracting text annotations of historical content includes the following steps:

[0056] The historical content is divided into multiple sub-segments based on the audio signal, and a multimodal model is constructed. Based on the multimodal model, multi-type data corresponding to the sub-segments are obtained, and the multi-type data includes video frames, background audio frames and text information.

[0057] Based on the multimodal model, the scene information of the video frame is obtained, the content label and the emotion label are set, and it is determined whether the semantic similarity between the scene information and the text information is greater than a second preset value. If so, the scene information and the text information are combined to generate fitting information, and the content categories and emotion categories corresponding to the content labels and the emotion labels are respectively extracted from the fitting information; otherwise, the content categories and emotion categories corresponding to the content labels and the emotion labels are respectively extracted from the text information and the scene information.

[0058] An audio database is established based on preset audio tags, reference audio with the greatest signal waveform similarity is extracted from the audio database based on the signal waveform of the background audio frame, and the preset audio tag corresponding to the reference audio is set as the audio category of the background audio frame.

[0059] The content categories, emotion categories and audio categories corresponding to all sub-segments are sequentially combined and set as text annotations of the historical content.

[0060] Specifically, in this embodiment, when the historical content is of video type, it may include background audio, motion image, voice audio and other contents at the same time. The audio signal refers to the audio signal of voice audio. The historical content can be divided into multiple sub-segments according to the fluctuation of the audio signal. The multimodal model refers to a machine learning or artificial intelligence model that can process and analyze multiple different types of data, including but not limited to machine vision algorithms and natural language processing algorithms generated by neural networks, etc., and can directly read multiple types of data of sub-segments, wherein the video frame refers to the motion image, the background audio frame refers to the background audio, and the text information refers to the text generated by the voice audio conversion. The scene information is used to describe the scene content of the video frame, including the scene, characters and actions of the motion image in the video frame. The content label refers to verbs and nouns, the emotion label refers to the emotion description word, and the semantic similarity refers to the meaning similarity between the two groups of phrases. The value of the semantic similarity can be obtained by the natural language processing algorithm in the multimodal model. If the semantic similarity is greater than the second preset value, it means that the content similarity between the scene information and the text information is high, the corresponding video frame and the meaning of the text information are consistent, and the video type corresponding to the sub-segment is consistent with the motion image and the text information. Fitting information refers to simple text generated by fitting scene information with text information, which can be generated by natural language processing model fitting. Content category refers to character information corresponding to content labels in fitting information, and emotion category refers to character information corresponding to emotion labels in fitting information. If the semantic similarity is less than or equal to the second preset value, it means that the meaning of the video frame and the text information is quite different, which can be simply understood as inconsistent graphic information. Therefore, it is necessary to extract content categories and emotion categories from both scene information and text information. Semantic similarity can be used to accurately obtain detailed information on historical content corresponding to the video type.

[0061] The preset audio tags are tags set according to different audio styles, including but not limited to light music, pop music, and classical music. The audio database is a collection of audios corresponding to multiple preset audio tags. Waveform similarity refers to the degree of fit between any two signal waveforms. The extracted reference audio has a high degree of similarity with the background audio frame. Therefore, the preset audio tag corresponding to the reference audio is set to the audio category of the background audio frame.

[0062] The content category, emotion category and audio category can comprehensively describe the changing rules and annotation summary of the historical content of the video type, that is, the text annotation.

[0063] As a preferred technical solution of the present invention, obtaining scene information of a video frame based on a multimodal model includes the following steps:

[0064] Collect video samples, mark the label indexes of the video samples, combine all video samples with the corresponding label indexes and set them as a training sample set, and generate a multimodal model based on the neural network and the feature vectors of the training sample set.

[0065] The multimodal model divides the video frame into multiple frame images based on preset frames, calculates the image similarity between the frame image and all video samples, generates a label index corresponding to the frame image based on the image similarity, sets it as the image information of the frame image, and sequentially combines all image information to generate scene information.

[0066] Specifically, in this embodiment, the video sample refers to a sample set of multiple types of video images, and the label index refers to the index summary information of the video sample, such as landscape, cultural landscape, dance movements, character explanations, etc. A multimodal model with high accuracy can be trained and generated by the training sample set and the neural network, wherein the feature vector of the training sample set refers to the feature extraction generated, for example, using Word2Vec (word embedding) to extract text features, using CNN (convolutional neural network) to extract image features, and using an acoustic model to extract audio features.

[0067] The preset frame refers to the preset frame value. Dividing the video frame into multiple frame images can facilitate the refined analysis of image features. The multimodal model generates a label index corresponding to the frame image based on the image similarity between the frame image and all video samples, that is, the image information. After combining the image information corresponding to all frame images in sequence, the scene information of the video frame can be generated, which can improve the integrity of the scene information.

[0068] As a preferred technical solution of the present invention, labeling the label index of the video sample includes the following steps:

[0069] All video samples are classified into labeled videos and test videos, relevant objects of the labeled videos are obtained based on image motion detection, keyword groups are generated based on the types of relevant objects, semantic tags corresponding to the keyword groups are generated based on a preset knowledge graph, the first probability of each semantic tag in the test video is output, semantic tags with a first probability greater than or equal to a third preset value are extracted and set as test tags, and the keyword groups and test tags are respectively marked as tag indexes corresponding to the video samples.

[0070] Specifically, in this embodiment, the marked video is a sample for comparison, the test video is a video for automatically generating a label index, image motion detection is the process of identifying the motion of feature vectors in multiple adjacent frames of an image using computer vision, and the related object refers to the image elements appearing in the marked video, including but not limited to people, scenery, and actions, etc. The keyword group is a character used to describe the related object. The knowledge graph is a structured semantic knowledge base that organizes complex knowledge in a visual and structured manner to facilitate computer program understanding, storage, and retrieval. The semantic label refers to the category label of the keyword group. The semantic label is a further subdivision of the concept and is used to distinguish different entity groups, such as "clouds" is a category of "scenery". The test video may contain multiple objects. The first probability refers to the frequency of the semantic label appearing in the test video, which can be output by the neural network. If the first probability is greater than or equal to the third preset value, the corresponding semantic label, i.e., the test label, is retained. By setting the keyword group as the label index corresponding to the marked video and setting the test label as the label index corresponding to the test video, the automatic labeling of the label index of the video sample can be accurately and quickly realized.

[0071] As a preferred technical solution of the present invention, generating a user's interest change curve based on a content change curve includes the following steps:

[0072] Construct a standard content tag set, calculate the text similarity between the index information and any standard content tag in the standard content tag set, normalize all text similarities to generate strength values ​​of the standard content tags, set all standard content tags as horizontal axis information, set the strength values ​​as vertical axis values, and generate a content change curve based on the horizontal axis information and the corresponding vertical axis values.

[0073] Arrange all the interesting contents in sequence based on the timestamps of the interaction data to generate a content time series, calculate the standard deviation corresponding to the content change curve of the interesting contents, set it as the value of the content time series at the timestamp, and connect the values ​​at all timestamps in sequence to generate an interest change curve.

[0074] Specifically, in this embodiment, the standard content tag set refers to a set of standard content tags, wherein the standard content tags include but are not limited to sports, positive, negative, adventure, and funny. Text similarity refers to the similarity of text characters, and the text similarity P can be calculated by the first formula, which is: , where m is the number of characters that are the same between the index information and the standard content label, and M is the total number of characters in the index information. Standardization processing refers to processing the text similarity standard into an interval The value in is the intensity value. A rectangular coordinate system is constructed, and the content change curve of the historical content can be generated by the horizontal axis information and the corresponding vertical axis value.

[0075] Timestamp refers to the time point corresponding to when the user browses historical content. All the content of interest is arranged in sequence based on the timestamp to generate a content time series. The standard deviation is a measure of the degree of deviation of the discrete data point set in the curve. The standard deviation corresponding to all the content of interest is set as the value corresponding to the content time series, and the user's interest change curve is generated in sequence. By analyzing the process of the user's content of interest changing over time, the recommended content corresponding to a specific time can be provided to the user more accurately.

[0076] As a preferred technical solution of the present invention, extracting the optimized content includes the following steps:

[0077] Obtain content change curves corresponding to all updated contents in the associated database and set them as the first curves, cluster all first curves to generate multiple cluster categories, set the updated contents contained in each cluster category as the first contents, calculate the average value of the standard deviations corresponding to all first contents, and use the standard recommendation model to output the predicted value of the predicted time point based on the interest change curve and timestamp, calculate the difference between the predicted value and the average value, set the first content corresponding to the minimum difference as the candidate content, and set the candidate content with the minimum standard deviation as the optimized content.

[0078] Specifically, in this embodiment, the updated content refers to the content data contained in the associated database, and clustering all the first curves can more accurately classify all the updated content into multiple cluster classifications based on relevance. The LSTM neural network in the standard recommendation model can obtain the changing characteristics of the user's interest change curve with the timestamp and predict it, and output the value of the interest change curve at the next timestamp, that is, the predicted value at the predicted time point. According to the predicted value, the alternative cluster classification can be located, and then based on the difference between the predicted value and the average value, a first content is extracted from all the first contents contained in the cluster classification, which is the alternative content. If there are multiple alternative contents with the same minimum difference, the alternative content with the minimum standard deviation is set as the optimized content. The predicted value of the standard recommendation model can accurately predict the user's browsing habits, and then accurately extract the optimized content and push it to the user.

[0079] As a preferred technical solution of the present invention, the change rate of each parameter in the optimization model is adjusted based on the change rate of the standard deviation of all the first contents to generate a standard recommendation model.

[0080] Specifically, in this embodiment, the standard recommendation model is generated by continuously adjusting the parameters of the optimization model, and has high prediction accuracy. When dynamically adjusting the parameters of the optimization model, the change rate of the standard deviation of all the first contents can be used as the basis for adjustment. For example, a cluster classification contains 10 first contents, and there are 10 values ​​corresponding to the standard deviations. The change rate of these 10 values ​​can first calculate the change rate between any two values, and then calculate the average of all the change rates, which is the change rate of 10 standard deviations. Finally, the value is set as the change rate of each parameter, and the optimization model is adjusted in turn to generate a standard recommendation model.

[0081] As a preferred technical solution of the present invention, extracting the summary text of the historical content includes the following steps:

[0082] Build a natural language processing model and obtain the text characters of the historical content based on the string type. The natural language processing model extracts semantic information from the text characters and obtains the character group with the highest word frequency in the text characters. The character group is combined with the semantic information to generate a summary text.

[0083] Specifically, in this embodiment, the Natural Language Processing (NLP) model is a type of computer algorithm and machine learning model specially designed to analyze, understand and generate human language. These models can process and understand natural language data and perform various language-related tasks. Text characters refer to the text recognized by the natural language processing model, semantic information refers to the semantic text output by the natural language processing model based on text characters, word frequency refers to the ratio between the number of times the same phrase appears in the text characters and the total number of all phrases in the text characters, character groups refer to the text corresponding to the maximum word frequency in the text characters, and combination refers to arranging character groups and semantic information based on word order vocabulary to generate summary text.

[0084] like Figure 2 As shown, the present invention also provides an AIGC content recommendation system based on user participation optimization, which is used to implement the above-mentioned AIGC content recommendation method based on user participation optimization. The system mainly includes:

[0085] The content collection module collects the interaction data between users and historical content, generates the user's engagement and the historical content's engagement based on the interaction data statistics, and filters out the user's content of interest from all historical content based on the engagement.

[0086] The content analysis module generates an information flow of historical content based on preset tags, extracts index information from the information flow based on preset descriptive words, generates a content change curve corresponding to the content of interest based on the index information, and generates a user's interest change curve based on the content change curve.

[0087] The content recommendation module builds an optimization model. The optimization model dynamically adjusts parameters within a preset time interval based on the content change curve and engagement to generate a standard recommendation model. The associated database is generated based on index information. The standard recommendation model extracts optimized content from the associated database based on the interest change curve and pushes it to users.

[0088] The model optimization module repeatedly executes S3 until the engagement degree between the user and the optimized content is greater than or equal to a first preset value, thereby completing the content recommendation.

[0089] It should be understood that, although each step in the flow chart of each embodiment of the present invention is shown in sequence according to the indication of the arrow, these steps are not necessarily performed in sequence according to the order indicated by the arrow. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.

[0090] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the above-mentioned program can be stored in a non-volatile computer-readable storage medium. When the 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. As an illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0091] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0092] The above embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for those of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

[0093] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An AIGC content recommendation method based on user participation optimization, characterized in that: The method comprises the following steps: S1: Collecting interaction data between the user and historical content, generating the user's engagement and the engagement of the historical content based on the interaction data, and filtering out the user's interested content from all historical content based on the engagement; S2: Generate an information stream of the historical content based on preset tags, extract index information from the information stream based on preset description words, generate a content change curve corresponding to the content of interest based on the index information, and generate an interest change curve of the user based on the content change curve; S3: constructing an optimization model, wherein the optimization model dynamically adjusts parameters within a preset time interval based on the content change curve and the engagement level to generate a standard recommendation model, and generates an associated database based on the index information, wherein the standard recommendation model extracts optimized content from the associated database based on the interest change curve and pushes the optimized content to the user; S4: Repeat S3 until the engagement level between the user and the optimized content is greater than or equal to a first preset value, thereby completing content recommendation.

2. The method according to claim 1, characterized in that The generating of the information flow of the historical content based on the preset tag comprises the following steps: The preset tags include string type, video type and image type. All the historical contents are classified into multiple subcategories based on the preset tags. If the subcategory is the string type, the summary text of the historical content is extracted. If the subcategory is the video type, the text annotation of the historical content is extracted. If the subcategory is the image type, the image tag of the historical content is extracted. The summary text, the text annotation and the image tag are respectively set as the information flow corresponding to the historical content.

3. The method according to claim 2, characterized in that The extracting of text annotations of the historical content comprises the following steps: Segmenting the historical content into a plurality of sub-segments based on an audio signal, constructing a multimodal model, and acquiring multi-type data corresponding to the sub-segments based on the multimodal model, wherein the multi-type data includes video frames, background audio frames, and text information; Based on the multimodal model, the scene information of the video frame is obtained, a content label and an emotion label are set, and it is determined whether the semantic similarity between the scene information and the text information is greater than a second preset value. If yes, the scene information and the text information are combined to generate fitting information, and the content category and the emotion category corresponding to the content label and the emotion label are respectively extracted from the fitting information; otherwise, the content category and the emotion category corresponding to the content label and the emotion label are respectively extracted from the text information and the scene information; Establish an audio database based on preset audio tags, extract reference audio with the greatest signal waveform similarity from the audio database based on the signal waveform of the background audio frame, and set the preset audio tag corresponding to the reference audio as the audio category of the background audio frame; The content categories, the emotion categories and the audio categories corresponding to all the sub-segments are sequentially combined and set as text annotations of the historical content.

4. The method according to claim 3, characterized in that The acquiring the scene information of the video frame based on the multimodal model comprises the following steps: Collecting video samples, marking label indexes of the video samples, combining all the video samples with the corresponding label indexes to form a training sample set, and generating the multimodal model based on a neural network and feature vectors of the training sample set; The multimodal model divides the video frame into multiple frame images based on a preset frame, calculates the image similarity between the frame image and all the video samples, generates a label index corresponding to the frame image based on the image similarity, sets it as the image information of the frame image, and sequentially combines all the image information to generate the scene information.

5. The method according to claim 4, characterized in that The label index of the video sample is marked, comprising the following steps: All the video samples are classified into labeled videos and test videos, related objects of the labeled videos are obtained based on image motion detection, keyword groups are generated based on the types of the related objects, semantic tags corresponding to the keyword groups are generated based on a preset knowledge graph, the first probability of each semantic tag in the test video is output, the semantic tag whose first probability is greater than or equal to a third preset value is extracted, set as a test tag, and the keyword group and the test tag are respectively marked as the tag index corresponding to the video sample.

6. The method according to claim 1, characterized in that Generating the user's interest change curve based on the content change curve comprises the following steps: Constructing a standard content tag set, calculating the text similarity between the index information and any standard content tag in the standard content tag set, normalizing all text similarities to generate strength values ​​of the standard content tags, setting all the standard content tags as horizontal axis information, setting the strength values ​​as vertical axis values, and generating the content change curve based on the horizontal axis information and the corresponding vertical axis values; Arrange all the contents of interest in sequence based on the timestamp of the interaction data to generate a content time series, calculate the standard deviation corresponding to the content change curve of the contents of interest, set it as the value of the content time series at the timestamp, connect all the values ​​at the timestamps in sequence, and generate the interest change curve.

7. The method according to claim 6, characterized in that The extracting optimization content comprises the following steps: Obtain content change curves corresponding to all updated contents in the associated database and set them as the first curve, cluster all the first curves to generate multiple cluster categories, set the updated contents contained in each cluster category as the first content, calculate the average value of the standard deviations corresponding to all first contents, the standard recommendation model outputs a predicted value at a predicted time point based on the interest change curve and the timestamp, calculates the difference between the predicted value and the average value, sets the first content corresponding to the minimum difference as the alternative content, and sets the alternative content with the minimum standard deviation as the optimized content.

8. The method according to claim 7, characterized in that The change rate of each parameter in the optimization model is adjusted based on the change rate of the standard deviation of all the first contents to generate the standard recommendation model.

9. The method according to claim 2, characterized in that: The extracting of the summary text of the historical content comprises the following steps: A natural language processing model is constructed to obtain text characters of the historical content based on the string type. The natural language processing model extracts semantic information from the text characters and obtains a character group with the highest word frequency in the text characters. The character group is combined with the semantic information to generate the summary text.

10. An AIGC content recommendation system based on user participation optimization, used to implement the method according to any one of claims 1 to 9, characterized in that: The system includes the following modules: A content collection module collects interaction data between a user and historical content, generates the user's engagement and the engagement of the historical content based on the interaction data, and selects the user's interested content from all historical content based on the engagement; A content analysis module, which generates an information stream of the historical content based on preset tags, extracts index information from the information stream based on preset descriptive words, generates a content change curve corresponding to the content of interest based on the index information, and generates an interest change curve of the user based on the content change curve; A content recommendation module, constructing an optimization model, wherein the optimization model dynamically adjusts parameters within a preset time interval based on the content change curve and the engagement level to generate a standard recommendation model, and generates an associated database based on the index information, wherein the standard recommendation model extracts optimized content from the associated database based on the interest change curve and pushes it to the user; The model optimization module repeatedly executes S3 until the engagement degree between the user and the optimized content is greater than or equal to a first preset value, thereby completing content recommendation.

Images