Television media asset data recommendation processing method and device, computer equipment and readable storage medium

Abstract

The invention discloses a television media resource data recommendation processing method and device, computer equipment and a readable storage medium, and the method comprises the steps: firstly, generating media resource vector data according to the vectorization processing of a local media resource library, and then obtaining a recommendation model through the training of a pre-training base model and a specific strategy; the method comprises the following steps: firstly, acquiring a user television operation behavior to generate a family portrait, and finally, according to the family portrait and media asset vector data, reasoning and fusing through a recommendation model to generate a final family user sheet recommendation result, thereby realizing accurate and personalized recommendation of the television media asset data.

Description

Technical Field

[0001] The present invention relates to the technical field of media resource recommendation, and in particular to a method, device, computer equipment and readable storage medium for recommending and processing television media resource data. Background Art

[0002] In the current TV media industry, how to provide users with accurate recommendations that meet family needs has always been a challenge. Existing recommendation methods are often not personalized enough, lack accuracy, and cannot fully consider the complex behaviors and diverse preferences of family users. At the same time, there are also inefficiencies in the processing of media data and the training of models, making it difficult to meet users' growing demand for high-quality viewing. Summary of the invention

[0003] The object of the present invention is to provide a method, device, computer equipment and readable storage medium for recommending and processing television media asset data.

[0004] In a first aspect, an embodiment of the present invention provides a method for recommending and processing television media asset data, including:

[0005] Perform media data vectorization processing based on the local media database to generate media vector data;

[0006] The pre-trained base model is used to train and fine-tune the preset training data set in combination with the low-rank adaptation strategy and the hybrid expert model to obtain a TV media data recommendation model.

[0007] Acquire the user's television operation behavior, and generate a family portrait according to the user's television operation behavior;

[0008] According to the family portrait and the media vector data, the recommendation result is generated by inference through the television media data recommendation model, and the final family user movie list recommendation result is generated by fusion using a fusion sorting algorithm.

[0009] In a possible implementation, the performing vectorization processing on the media asset data according to the local media asset library to generate the media asset vector data includes:

[0010] Acquire media asset data in the local media asset library, wherein the media asset data includes a media asset ID, a media asset name, a media asset introduction, a media asset plot, basic tags, a region, and a release time;

[0011] Preprocessing the media asset data to convert independent field information of the media asset data into a natural language description;

[0012] Perform word segmentation processing on the natural language description to obtain a word segmentation list;

[0013] The word segmentation list is vectorized using the FastText model to obtain the media asset vector data.

[0014] In a possible implementation, the pre-trained base model is used to train and fine-tune a preset training data set in combination with a low-rank adaptation strategy and a hybrid expert model to obtain a TV media data recommendation model, including:

[0015] Using the llama38B model as the pre-trained base model;

[0016] The preset training data set is trained and fine-tuned according to the llama38B model combined with the low-rank adaptation strategy and the hybrid expert model to obtain a television media data recommendation model.

[0017] In a possible implementation manner, the acquiring of the user's television operation behavior and generating a family portrait according to the user's television operation behavior includes:

[0018] Acquire the user's TV operation behavior within a preset time range, the user's TV operation behavior includes user play behavior and user click behavior, and the media asset ID, media asset name, time point, play duration, and weekend and holiday identifier of the media asset data corresponding to the user play behavior and the click behavior;

[0019] Preprocessing the user's television operation behavior to obtain a user behavior summary description based on a natural language style corresponding to the user's television operation behavior;

[0020] The family portrait is obtained according to the summary description of the user behavior.

[0021] In a possible implementation manner, the user behavior summary description includes a weekday behavior summary description and a weekend and holiday behavior summary description;

[0022] The step of obtaining the family portrait according to the user behavior summary description includes:

[0023] The weekday behavior summary description and the weekend and holiday behavior summary description are based on the custom large model prompt words, and the family member distribution and viewing time period preference of the TV user are inferred to obtain the weekday family portrait and the weekend and holiday family portrait.

[0024] In a possible implementation, the generating of recommendation results by inference of the television media data recommendation model based on the family portrait and the media vector data, and the generating of the final family user movie list recommendation results by fusing with a fusion sorting algorithm include:

[0025] Get current time information;

[0026] Determine a weekday family portrait or a weekend or holiday family portrait according to the current time information;

[0027] According to the media vector data and the weekday family portrait or weekend and holiday family portrait, the recommendation result is generated by reasoning through the television media data recommendation model, and the final family user movie list recommendation result is generated by fusion using a fusion sorting algorithm.

[0028] In a possible implementation, the generating of recommendation results by inference of the TV media data recommendation model based on the media asset vector data and the weekday family portrait or the weekend and holiday family portrait, and generating the final family user movie list recommendation result by fusing with a fusion sorting algorithm, includes:

[0029] Generate a first recommendation result by inferring the television media data recommendation model according to the media asset vector data and the weekday family portrait or the weekend and holiday family portrait;

[0030] Generate a second recommendation result based on the weekday family portrait or the weekend and holiday family portrait based on text similarity and collaborative filtering algorithms;

[0031] The first recommendation result and the second recommendation result are fused by using a fusion sorting algorithm to obtain the final home user movie list recommendation result.

[0032] In a second aspect, an embodiment of the present invention provides a device for recommending and processing television media asset data, including:

[0033] The acquisition module is used to perform vector processing on the media data according to the local media library to generate media vector data;

[0034] A training module is used to train and fine-tune a preset training data set using a pre-trained base model in combination with a low-rank adaptation strategy and a hybrid expert model to obtain a TV media data recommendation model;

[0035] The recommendation module is used to obtain the user's TV operation behavior and generate a family portrait based on the user's TV operation behavior; based on the family portrait and the media vector data, generate recommendation results through reasoning through the TV media data recommendation model, and use the fusion sorting algorithm to generate the final family user movie list recommendation results.

[0036] In a third aspect, an embodiment of the present invention provides a computer device, comprising a processor and a non-volatile memory storing computer instructions, wherein when the computer instructions are executed by the processor, the computer device executes the method described in any one of at least one possible implementation of the first aspect.

[0037] In a fourth aspect, an embodiment of the present invention provides a readable storage medium, wherein the readable storage medium includes a computer program, and when the computer program is executed, the computer device where the readable storage medium is located is controlled to execute the method described in at least one possible implementation manner of the first aspect.

[0038] Compared with the prior art, the beneficial effects provided by the present invention include: adopting a recommendation processing method, device, computer equipment and readable storage medium for television media data disclosed in the present invention, generating media vector data through vectorization processing according to a local media library, and then using a pre-trained base model combined with specific strategy training to obtain a recommendation model, and then obtaining the user's television operation behavior to generate a family portrait, and finally, based on the family portrait and media vector data, the recommendation model is inferred and integrated to generate the final family user film list recommendation result, thereby achieving accurate and personalized recommendation of television media data. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative work.

[0040] Figure 1 A schematic diagram of the steps of a method for recommending processing television media asset data provided by an embodiment of the present invention;

[0041] Figure 2 A schematic diagram of a framework of a recommended processing flow for television media asset data provided by an embodiment of the present invention;

[0042] Figure 3 A schematic block diagram of the structure of a device for recommending and processing television media asset data provided by an embodiment of the present invention;

[0043] Figure 4 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0045] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.

[0046] In order to solve the technical problems in the aforementioned background technology, Figure 1 This is a flow chart of a method for recommending processing television media asset data provided by an embodiment of the present disclosure. The method for recommending processing television media asset data is introduced in detail below.

[0047] Step S201, performing vectorization processing on media data according to the local media database to generate media vector data;

[0048] Step S202, using the pre-trained base model, combined with the low-rank adaptation strategy and the hybrid expert model, to train and fine-tune the preset training data set to obtain a TV media data recommendation model;

[0049] Step S203, obtaining the user's television operation behavior, and generating a family portrait according to the user's television operation behavior;

[0050] Step S204, based on the family portrait and the media vector data, the recommendation result is generated by inference through the television media data recommendation model, and the final family user movie list recommendation result is generated by fusion sorting algorithm.

[0051] In the embodiment of the present invention, for example, it is assumed that our server operates a platform with rich television media resources, and a large amount of various program contents are stored in the local media resource library.

[0052] First, the server obtains the media data in the local media library. This data may contain various information. For example, the movie "Interstellar XX" with media ID 001 has a media name of "Interstellar XX", a media introduction of "a science fiction adventure movie about humans traveling through wormholes to distant galaxies to find a new home", and a media plot of "Ku X is a former astronaut who discovered a strange gravitational phenomenon in his daughter Mo X's study, and thus embarked on a journey through a wormhole to find a new home. In the process, he faced many challenges such as time dilation and black hole gravity, while also constantly pursuing the hope of saving mankind", and the basic tags are "science fiction, adventure, time travel", and the release time is "2014".

[0053] Next, the server preprocesses these media data. For the above-mentioned movie "Interstellar XX", the independent field information is converted into natural language descriptions. For example, labels such as "science fiction, adventure, time travel" are converted into "This is a movie with science fiction elements, full of adventure plots, and involving the theme of time travel." The release date "2014" is converted into "This movie was released in 2014." After such processing, the originally discrete independent field information is integrated into a more coherent natural language description that is easier to process later.

[0054] Then, the server performs word segmentation on these natural language descriptions. For the natural language description of "Interstellar XX", the following word segmentation list may be obtained: "a", "about", "human", "crossing", "wormhole", "going to", "distant", "galaxy", "looking for", "new home", "of", "science fiction", "adventure", "movie", "KuX", "is", "one", "former", "astronaut", "he", "in", "daughter", "MoX", "of", "study", "discover", "strange", "of", "gravity", "phenomenon", "thus", "set foot", "of", "crossing", "wormhole", "looking for", "new home", "of", "journey", "in", "this", "process", "in", "he", "facing", "with", "time", "expansion", "black hole", "gravity", "etc.", "many", "challenges", "at the same time", "also", "in", "continuously", "pursuing", "with", "saving", "human", "of", "hope", "this is", "a", "with", "science fiction", "elements", "full of", "adventure", "plot", "involving", "time and space", "crossing", "theme", "of", "movie", "this", "movie", "from", "this", "movie", "in", "2014", "released".

[0055] Finally, the server uses the FastText model to vectorize these word lists. The FastText model generates a corresponding vector for each word, and then combines or averages these word vectors to finally obtain the media vector data of the movie "Interstellar XX". This vector data can represent various features and information of the movie in numerical form, providing a basic quantitative representation for subsequent recommendation processing.

[0056] By performing the above processing on all media data in the local media library, the server can generate a large amount of media vector data, thereby providing rich input for subsequent recommendation models.

[0057] The server selects the llama38B model as the pre-trained base model. The preset training data set contains a large amount of user behavior data, media data, and corresponding recommendation result labels.

[0058] First, the server inputs the preset training data set into the llama38B model. The model begins to learn and understand the data, trying to capture the potential patterns and rules in the data.

[0059] During the training process, a low-rank adaptation strategy is combined. Assume that a portion of users in the training data set have a preference for science fiction movies that shows low-dimensional characteristics, for example, they pay more attention to the special effects scenes and future settings in the movies. The low-rank adaptation strategy can identify this low-dimensional feature pattern and adjust the parameters of the model in a targeted manner to better adapt to the preferences of such users.

[0060] At the same time, a hybrid expert model is used. Assume that there are different types of users in the data set, some of whom are action movie fans, while others prefer romance movies. The hybrid expert model can configure dedicated "expert" modules for different types of users. Each expert module focuses on processing the preference data of a specific type of user, thereby improving the recommendation accuracy and personalization of the model.

[0061] After continuous iterative training and parameter adjustment, the server fine-tunes the model. For example, if it is found that the model does not recommend certain popular movies that have just been released, the server will specifically add data related to these movies for retraining to optimize the model's performance.

[0062] After a long period of training and fine-tuning, the server finally obtained a TV media data recommendation model that can accurately predict users' preferences for TV media. This model can generate accurate recommendation results for users based on their historical behavior and the characteristics of the media.

[0063] The server continuously monitors the user's TV operation behavior. Assume that in a family, the following user TV operation behaviors were recorded in the past week:

[0064] User A watched the movie "XX" from 8 to 10 p.m. on Monday, which lasted a full two hours.

[0065] User B watched some episodes of the TV series "XX List" from 3:00 to 4:30 p.m. on Wednesday.

[0066] User C watched the variety show "Running Man" from 7 to 9 pm on Saturday.

[0067] User A and User B watched the documentary "Earth XX" together from 2:00 pm to 5:00 pm on Sunday.

[0068] The server first obtains the user's TV operation behavior within the preset time range (one week), including user playback behavior and user click behavior, as well as detailed information such as the media asset ID, media asset name, time point, playback duration, weekend and holiday identification, etc. of the media asset data corresponding to these behaviors.

[0069] Then, these user TV operation behaviors are preprocessed. The above behaviors are converted into a summary description of user behaviors based on natural language style, for example: "User A started watching the movie "XX Story" at 8 pm on Monday, which lasted for two hours. User B started watching the TV series "XX List" at 3 pm on Wednesday, which lasted for one and a half hours. User C started watching the variety show "Running XX" at 7 pm on Saturday, which lasted for two hours. User A and user B started watching the documentary "Earth XX" together at 2 pm on Sunday, which lasted for three hours."

[0070] Based on these user behavior summary descriptions, the server further analyzes and obtains family portraits. The user behavior summary descriptions include weekday behavior summary descriptions (such as user A's viewing behavior on Monday and user B's viewing behavior on Wednesday) and weekend and holiday behavior summary descriptions (such as user C's viewing behavior on Saturday and users A and B's viewing behavior on Sunday).

[0071] For this family, the server uses the summary description of weekday behavior and weekend and holiday behavior based on the custom large model prompt words to infer the distribution of TV users' family members and viewing time preferences. For example, through analysis, it is found that most of the content watched on weekday evenings is drama-based content, while on weekends, more entertainment-based programs are preferred. As a result, the weekday family portrait of this family is "family members A and B have the habit of watching drama-based films and TV shows on weekday evenings, and the viewing time is longer", and the weekend and holiday family portrait is "family members A, B, and C are more inclined to watch entertainment-based programs on weekends, and the viewing time is more concentrated in the afternoon and evening."

[0072] Assume that the current time is 8:00 p.m. on Saturday, and the server obtains the current time information.

[0073] Based on the current time, the server determines that the weekend and holiday family portrait should be used. Then, the server uses the TV media data recommendation model to make inferences based on the previously generated media vector data (such as the vector data of the movie "XX Detective" and the vector data of the TV series "Secret XX") and the weekend and holiday family portrait.

[0074] The model generates the first recommendation result based on the preference for entertainment programs in the family portrait and the characteristics in the media vector data. For example, it recommends the comedy movie "XX Richest Man" and the parent-child variety show "Where is XX Going".

[0075] At the same time, the server generates the second recommendation result based on the family portrait on weekends and holidays, based on text similarity and collaborative filtering algorithms. Assuming that this family often watched comedy-related programs in the past, based on the text similarity algorithm, it will recommend a similar comedy movie "XX Iron Fist"; based on the collaborative filtering algorithm, it will find that other families with similar viewing behaviors like to watch the parent-child reality show "XX is Superman", which will also be included in the recommendation.

[0076] Next, the server uses a fusion sorting algorithm to merge the first recommendation result and the second recommendation result. The fusion sorting algorithm takes into account multiple factors, such as the user's historical viewing preferences, the popularity of the recommendation results, and the program update time. For the above example, a comprehensive sorting may be performed based on the popularity of the program and the degree of match with the family portrait. Finally, the final family user movie list recommendation result generated for this family may be arranged in the following order: "XX Richest Man", "XX Iron Fist", "XX Where to Go", and "XX is Superman".

[0077] In this way, the server generates a personalized and accurate list of recommended TV media resources for the family during the current weekend and holidays, which can meet the viewing needs and preferences of family members and enhance the user's viewing experience.

[0078] In the embodiment of the present invention, the aforementioned step S201 can be implemented through the following example execution.

[0079] Acquire media asset data in the local media asset library, wherein the media asset data includes a media asset ID, a media asset name, a media asset introduction, a media asset plot, basic tags, a region, and a release time;

[0080] Preprocessing the media asset data to convert independent field information of the media asset data into a natural language description;

[0081] Perform word segmentation processing on the natural language description to obtain a word segmentation list;

[0082] The word segmentation list is vectorized using the FastText model to obtain the media asset vector data.

[0083] In the embodiment of the present invention, illustratively, the server first obtains the media data in the local media library. For example, a TV series "XX Side Story" with a media ID of 002 is obtained, and its media name is "XX Side Story", and the media introduction is "An inn team headed by the shopkeeper Tong XX, performed classic funny scenes under the same roof, and accompanied the audience to grow up gradually in laughter and tears", and the media plot is "The story revolves around a female shopkeeper Tong XX and her several assistants in the XX Inn in the XX town in the virtual Ming Dynasty (approximately during the Wanli period). This group of young people performed classic funny scenes under the same roof, and accompanied the audience to grow up gradually in laughter and tears", the basic tags are "comedy, costume, situation comedy", and the release time is "2006".

[0084] Next, the server preprocesses these media data. The independent field information of the media data is converted into natural language descriptions. For example, for the basic tags "comedy, ancient costume, situation comedy", it is converted into "This is a TV series with comedy elements, set in ancient times, and belongs to the situation comedy type"; for the release time "2006", it is converted into "This TV series was released in 2006". After such processing, the originally scattered independent field information is integrated into a smoother and more logical natural language description, that is, "An inn team headed by the shopkeeper Tong XX, performed classic funny scenes under the same roof, and accompanied the audience to grow up together in laughter and tears. This is a TV series with comedy elements, set in ancient times, and belongs to the situation comedy type. It was released in 2006."

[0085] Then, the server performs word segmentation on the natural language description. The resulting word segmentation list may be "with" "shopkeeper", "Tong XX", "headed", "of", "an", "inn", "team", "in", "same", "under the eaves", "interpreted", "scene after scene", "classic", "of", "funny", "scene", "in", "laughter", "with", "tears", "in", "accompany", "audience", "together", "gradually", "grow up", "this is", "a", "with", "comedy", "elements", "background", "setting", "in", "ancient times", "belonging to", "situation comedy", "type", "of", "TV series", "released" in "2006".

[0086] Finally, the server uses the FastText model to vectorize the word list. The FastText model generates a corresponding vector representation for each word. For example, the word "shopkeeper" may be represented as a vector of a specific dimension [0.1, 0.2, 0.3, ...], and the word "comedy" may be represented as another vector [0.5, 0.6, 0.7, ...]. Then, the vectors of all the word segments are combined or calculated in some way (such as average) to finally obtain the media asset vector data of the TV series "XX Side Story". This vector data can accurately capture and express various characteristics and information of the TV series in a digital form.

[0087] The media vector data of the movie is obtained by vectorization using the FastText model. The server generates rich and accurate media vector data for subsequent recommendation work by processing a large amount of media data in the local media library in this way.

[0088] In the embodiment of the present invention, the aforementioned step S202 may be implemented through the following example execution.

[0089] Using the llama38B model as the pre-trained base model;

[0090] The preset training data set is trained and fine-tuned according to the llama38B model combined with the low-rank adaptation strategy and the hybrid expert model to obtain a television media data recommendation model.

[0091] In the embodiment of the present invention, the server first determines to use the llama38B model as the pre-training base model. This model has powerful language understanding and generation capabilities and can process large-scale text data.

[0092] The server obtains a preset training data set, which contains a large amount of interaction data between users and TV media. For example, there is a list of TV series watched by user A in the past month, including the time, duration, and episodes watched repeatedly; user B's ratings and comments on different types of movies; user C's search and click behavior on specific variety shows, etc.

[0093] At the beginning of training, the server processes the preset training data set according to the llama38B model combined with the low-rank adaptation strategy. Assume that in the data set, it is found that some users show a specific pattern of preference for costume dramas. They prefer to watch costume dramas with historical depth, exquisite costumes and props, and complex character relationships. The low-rank adaptation strategy can identify this low-dimensional but significant feature pattern. The server adjusts the parameters of the model so that the model can more keenly capture the special preference of such users for costume dramas, thereby improving the accuracy of recommending costume drama-related media resources to these users.

[0094] At the same time, the server uses a hybrid expert model. Assume that user groups of different age groups can be distinguished in the data set. For example, young users prefer to watch anime and youth idol dramas, middle-aged users are more interested in family ethics dramas and suspense dramas, and elderly users may prefer operas and classic historical dramas. The hybrid expert model configures special "expert" modules for these different user groups. The expert module for young users will focus on learning and analyzing data features related to anime and youth idol dramas, the expert module for middle-aged users will focus on processing preference data for family ethics dramas and suspense dramas, and the expert module for elderly users will focus on understanding the characteristics of operas and historical dramas. In this way, the model can provide more accurate and more in-line recommendations for users of different ages.

[0095] During the training process, the server continuously adjusts the parameters of the model. For example, if it is found that the model's recommendation effect on some newly launched popular TV series is not ideal, the server will increase the amount of data related to these TV series and retrain. For another example, if the model does not accurately grasp the long-term viewing preferences of some users, the server will conduct a more in-depth analysis of the data of these users and adjust the model.

[0096] After multiple iterations of training and fine-tuning, the server finally obtained a performance-optimized TV media data recommendation model that can accurately predict users’ interests and preferences for different TV media based on their complex and diverse behavioral data.

[0097] For example, when a new user D logs in to the platform, the model can quickly analyze and recommend other related TV series that may suit his interests based on D's initial operation behavior, such as the types of TV series he browsed and the length of time he stayed on the platform. Or when user E's viewing preferences change, for example, from mainly watching comedies in the past to paying more attention to science fiction themes recently, the model can capture this change in time and adjust the recommended content to recommend the latest and most popular science fiction films and TV shows to E.

[0098] For another example, for a family user group, which includes members of different ages and genders, the server analyzes their common viewing behaviors and their respective special preferences, and uses the trained recommendation model to recommend warm family dramas suitable for the whole family to watch together, and provide personalized recommendations for each member, such as recommending interesting cartoons for children, exciting suspense dramas for adults, and classic opera programs for the elderly.

[0099] By continuously utilizing large amounts of real user data for training and fine-tuning, the server ensures that the recommendation model can continue to learn and improve, in order to provide more accurate and thoughtful TV media recommendation services to meet users' increasingly diverse and personalized viewing needs.

[0100] In the embodiment of the present invention, the aforementioned step S203 can be implemented through the following example execution.

[0101] Acquire the user's TV operation behavior within a preset time range, the user's TV operation behavior includes user play behavior and user click behavior, and the media asset ID, media asset name, time point, play duration, and weekend and holiday identifier of the media asset data corresponding to the user play behavior and the click behavior;

[0102] Preprocessing the user's television operation behavior to obtain a user behavior summary description based on a natural language style corresponding to the user's television operation behavior;

[0103] The family portrait is obtained according to the summary description of the user behavior.

[0104] In the embodiment of the present invention, the server sets the preset time range to one month and starts to obtain the user's TV operation behavior during this period. For example, the following series of behaviors are recorded:

[0105] User playback behavior: The user started playing the movie "XX Animals" with media ID 001 at 8:00 p.m. on July 5, which was Sunday. The playback time was 1 hour and 30 minutes.

[0106] User click behavior: The user clicked on the TV series "XX List" with media ID 002 at 2 pm on July 12, but it was not played. That day was Tuesday.

[0107] User playback behavior: The user started playing the variety show "XX Life" with media ID 003 at 7:00 p.m. on July 18, which was a Saturday. The playback duration was 2 hours.

[0108] User playback behavior: The user started playing the documentary "Earth XX" with media ID 004 at 10 a.m. on July 25, which was Sunday. The playback time was 45 minutes.

[0109] The server obtains these user TV operation behaviors, including user playback behaviors and user click behaviors, as well as detailed information such as the media asset ID, media asset name, time point, playback duration, weekend and holiday identification, etc. of the media asset data corresponding to these behaviors.

[0110] Next, the server pre-processes these user TV operation behaviors and converts the above behaviors into a summary description of user behaviors based on natural language style. For example:

[0111] "At 8pm on July 5, the user started playing the movie "XX Animals" for 1 hour and 30 minutes. That day was Sunday. At 2pm on July 12, the user clicked on the TV series "XX List" but did not play it. That day was Tuesday. At 7pm on July 18, the user started playing the variety show "XX Life" for 2 hours. That day was Saturday. At 10am on July 25, the user started playing the documentary "Earth XX" for 45 minutes. That day was Sunday."

[0112] The server then summarizes the descriptions of these user behaviors to derive a family portrait.

[0113] Through analysis, it was found that on weekends (Sundays), users are more inclined to watch light-hearted and entertaining movies and variety shows, such as "XX Animals" and "XX Life", and the playback time is longer; while on weekdays (Tuesdays), users are more likely to click on but not play TV series, probably because they do not have enough time to watch due to work or study.

[0114] Based on this analysis, the server draws some characteristics of this family portrait: family members have more leisure time on weekends and like to watch programs that bring them joy and relaxation; family members are busier on weekdays and have limited time to watch TV, so they may be more inclined to choose content that can provide information or light entertainment in a short period of time.

[0115] By continuously acquiring and analyzing the TV operation behaviors of a large number of households, the server can continuously improve and refine the family portrait, thereby providing each family with more accurate and characteristic TV media recommendations. For example, for the first family mentioned above, more popular entertainment variety shows or comedy movies may be recommended on subsequent weekends; for the second family, other family ethics dramas of the same type may be recommended on weekday evenings, and movies suitable for parents and children to watch together may be recommended on weekends.

[0116] The server will also consider more factors to enrich the family portrait, such as differences in operating behaviors of different family members, usage of viewing devices, whether there is repeated viewing behavior, etc. Assuming that in the first family mentioned above, it is found that the same comedy movie is played by different user accounts many times on weekends, the server will think that this family has a high degree of liking for this type of comedy movie and give it a higher weight in subsequent recommendations.

[0117] For example, in the second family, if it is found that the child account frequently watches several specific cartoons, the server will focus on recommending works in the same series of these cartoons or other cartoons of similar styles when recommending children-related content for this family.

[0118] In short, by comprehensively and meticulously acquiring and analyzing users' TV operation behaviors, the server can generate accurate, rich and practical family portraits, providing each family with personalized and considerate TV media services.

[0119] In an embodiment of the present invention, the user behavior summary description includes a weekday behavior summary description and a weekend and holiday behavior summary description; the above-mentioned obtaining the family portrait based on the user behavior summary description can be implemented through the following example.

[0120] The weekday behavior summary description and the weekend and holiday behavior summary description are based on the custom large model prompt words, and the family member distribution and viewing time period preference of the TV user are inferred to obtain the weekday family portrait and the weekend and holiday family portrait.

[0121] In the embodiment of the present invention, exemplarily, the user behavior summary description obtained by the server is clearly divided into a weekday behavior summary description and a weekend and holiday behavior summary description. For example, for a family, its weekday behavior summary description is: "At 7 o'clock on Monday evening, the user played the news program with media ID 010 for 30 minutes. At 8 o'clock on Tuesday evening, the user clicked on the workplace drama with media ID 011, but it was not played. At 9 o'clock on Wednesday evening, the user played the knowledge lecture with media ID 012 for 1 hour. At 10 o'clock on Thursday evening, the user played the documentary with media ID 013 for 45 minutes. At 8:30 on Friday evening, the user played the historical drama with media ID 014 for 1 hour and 30 minutes."

[0122] The weekend and holiday behavior summary description is: "At 10 a.m. on Saturday, the user played the parent-child cartoon with media ID 015 for 1 hour and 15 minutes. At 2 p.m. on Saturday, the user played the comedy movie with media ID 016 for 2 hours. At 7 p.m. on Sunday, the user played the music variety show with media ID 017 for 2 hours and 30 minutes."

[0123] The server processes these descriptions based on the custom big model prompt words. The custom big model prompt words may include some specific questions and analysis directions, such as: "What time period do family members mainly watch TV on weekdays? What types of programs do they watch mainly? What are the characteristics of the viewing time distribution of family members on weekends and holidays? What types of programs do they prefer?"

[0124] Through comprehensive analysis of these prompt words and summary descriptions of user behavior, the server infers and generates the distribution of TV users' family members and their preferences for viewing time periods.

[0125] For this family, the server analysis shows that on weekdays, family members mainly watch TV between 7pm and 10pm, and tend to watch informational programs, such as news programs, knowledge lectures and documentaries, as well as work-related or in-depth dramas, such as historical dramas and workplace dramas. This shows that family members may be busy on weekdays, and they watch these programs to relax while also hoping to gain useful knowledge and information.

[0126] During weekends and holidays, family members watch movies at a more dispersed time, with movies being watched in the morning, afternoon and evening. They prefer programs that are relaxing and entertaining and suitable for the whole family to watch together, such as parent-child cartoons, comedy movies and music variety shows. This reflects that family members have more leisure time during weekends and holidays, and hope to enjoy family time, pleasure and relaxation by watching these programs.

[0127] Based on the above reasoning results, the server obtains the weekday family portrait and weekend and holiday family portrait of this family.

[0128] The weekday family portrait is: "Family members have the habit of watching TV on weekday evenings, mainly focusing on programs that provide knowledge and information, as well as dramas with a certain depth. The viewing time is relatively short and relatively fixed."

[0129] The family portrait on weekends and holidays is: "Family members are more flexible in watching movies on weekends and holidays, preferring programs that are relaxing and entertaining and suitable for the whole family to share, and watching movies for a longer time."

[0130] For another example, the summary of weekday behavior of another family is described as follows: "At 6:30 p.m. on Monday, the user played a fitness teaching video with media ID 018 for 40 minutes. At 9 p.m. on Tuesday, the user played a foreign language learning course with media ID 019 for 1 hour and 20 minutes. At 7 p.m. on Wednesday, the user clicked on a suspense drama with media ID 020, but it was not played. At 8:30 p.m. on Thursday, the user played a financial analysis program with media ID 021 for 1 hour and 10 minutes. At 9:30 p.m. on Friday, the user played a motivational speech with media ID 022 for 50 minutes."

[0131] The weekend and holiday behavior summary description is: "At 3 pm on Saturday, the user played the costume idol drama with media ID 023 for 2 hours and 20 minutes. At 8 pm on Saturday, the user played the family gathering game show with media ID 024 for 2 hours and 15 minutes. At 11 am on Sunday, the user played the travel documentary with media ID 025 for 1 hour and 40 minutes."

[0132] The server analyzes the custom large model prompts and finds that this family pays more attention to self-improvement and knowledge acquisition on weekdays, and their movie-watching time is relatively short and concentrated between 6:30 and 9:30 p.m. On weekends and holidays, their movie-watching time is more dispersed, and they prefer more entertaining programs that can bring joy and relaxation, such as costume idol dramas, family gathering game shows, and travel documentaries.

[0133] The resulting weekday family portrait for this family is: "Family members devote themselves to self-improvement on weekday evenings, watching programs related to study, work, and personal development, and the viewing time is compact."

[0134] The portrait of a family on weekends and holidays is: "Family members enjoy leisure time on weekends and holidays, have flexible viewing times, prefer entertainment and relaxation programs, and pursue a pleasant family atmosphere."

[0135] By conducting such in-depth and detailed analysis of the aggregated descriptions of user behavior of a large number of households, the server can generate accurate and targeted weekday family portraits and weekend and holiday family portraits for each household, thus laying a solid foundation for the subsequent provision of more accurate and family-friendly TV media recommendation services.

[0136] In the embodiment of the present invention, the aforementioned step S204 may be implemented through the following example execution.

[0137] Get current time information;

[0138] Determine a weekday family portrait or a weekend or holiday family portrait according to the current time information;

[0139] According to the media vector data and the weekday family portrait or weekend and holiday family portrait, the recommendation result is generated by reasoning through the television media data recommendation model, and the final family user movie list recommendation result is generated by fusion using a fusion sorting algorithm.

[0140] In the embodiment of the present invention, illustratively, the server first obtains current time information, assuming that the current time is 9 o'clock on Saturday night.

[0141] Based on this current time information, the server determines that the weekend and holiday family portrait should be used.

[0142] Then, the server obtains the previously generated media asset vector data. For example, the media assets corresponding to the media asset vector data include the popular movie "XX Detective 3", the variety show "XX Ace", the TV series "Secret XX", etc.

[0143] Next, the server uses the TV media data recommendation model to make inferences based on the weekend and holiday family portrait and these media vector data. Assume that the weekend and holiday family portrait of this family shows that they like entertainment programs that are relaxing, humorous, and have celebrities. Then, the recommendation model will analyze these features and media vector data.

[0144] For the movie "XX Detective 3", the model will consider the match between its comedy elements, star lineup, etc. and the family portrait, and believe that this movie is in line with the family's viewing preferences on weekends, thus generating recommendations.

[0145] The variety show "XX Ace" was also judged by the model to be suitable for recommendation because of its light-hearted and funny segments and many well-known stars.

[0146] For the TV series "Secret XX", although it is a high-quality drama, the model may not give it priority recommendation because its plot is relatively tense and suspenseful, which does not fit the family's pursuit of a relaxing and entertaining atmosphere on weekends.

[0147] After generating these recommendation results, the server uses a fusion sorting algorithm to integrate them. The fusion sorting algorithm takes into account multiple factors, such as the popularity of the media resource, the degree of matching with the family portrait, and the user's historical viewing history.

[0148] Assuming that "XX Detective 3" is a very popular movie at present and has a high degree of match with the family portrait, it will be given a higher weight in the ranking. Although "XX Ace" also fits the family portrait, it is less popular and has a relatively lower weight.

[0149] After being processed by the fusion sorting algorithm, the final recommended result for the family user's movie list for this family may be: "XX Detective 3" ranks first, followed by "XX Ace".

[0150] Let's assume another scenario, the current time is 8 o'clock on Wednesday night.

[0151] After the server obtains the current time, it determines to use the weekday family portrait.

[0152] The media vector data obtained include the knowledge lecture "World History XX", the workplace drama "Ideal XX", the documentary "The Rise of XX", etc.

[0153] According to the weekday family portrait, for example, this family is more inclined to acquire knowledge and work-related content on weekdays. The recommendation model infers that the knowledge lecture "World History XX" and the workplace drama "Ideal XX" match the portrait.

[0154] Although the documentary "The Rise of XX" is somewhat informative, it may not be direct and practical enough to meet the needs of this family on weekdays.

[0155] Using the fusion sorting algorithm and taking various factors into consideration, the final film list recommendation result may be: the workplace drama "Ideal XX" ranks first, and the knowledge lecture "World History XX" ranks second.

[0156] In this way, the server can generate accurate and appropriate film list recommendations for home users based on different times and corresponding family portraits, combined with media vector data, thereby improving users' viewing experience and satisfaction.

[0157] In an embodiment of the present invention, the aforementioned method of generating recommendation results based on the media vector data and the weekday family portrait or the weekend and holiday family portrait is performed by inferring the television media data recommendation model, and generating the final family user film list recommendation results by using a fusion sorting algorithm. This can be implemented through the following examples.

[0158] Generate a first recommendation result by inferring the television media data recommendation model according to the media asset vector data and the weekday family portrait or the weekend and holiday family portrait;

[0159] Generate a second recommendation result based on the weekday family portrait or the weekend and holiday family portrait based on text similarity and collaborative filtering algorithms;

[0160] The first recommendation result and the second recommendation result are fused by using a fusion sorting algorithm to obtain the final home user movie list recommendation result.

[0161] In the embodiment of the present invention, exemplarily, the server first generates a first recommendation result through inference of a television media data recommendation model based on the media vector data and the family portrait.

[0162] Assuming that it is the weekend, the obtained weekend and holiday family portrait shows that this family likes to watch parent-child programs and comedy movies. The media vector data includes the movie "XX Daddy", the variety show "Where is XX Going", the animated movie "XX Animals", etc. The TV media data recommendation model analyzes and infers the family portrait and media vector data, and believes that "XX Daddy" and "Where is XX Going" are very consistent with the weekend preferences of this family, and generates the first recommendation result.

[0163] Next, the server generates the second recommendation result based on the weekend and holiday family portraits, text similarity and collaborative filtering algorithms.

[0164] Based on the text similarity algorithm, since this family is interested in parent-child programs, the server will search for media assets with similar text descriptions related to "parent-child". For example, it is found that the TV series "Home with XX" has similarities in plot and theme with the parent-child programs that have been watched, and it is included in the recommendation.

[0165] Through the collaborative filtering algorithm, the server discovered that other users with similar weekend and holiday family portraits often watched the comedy animated film "XX Panda", and judged that this family might also be interested, so it was recommended as the second result.

[0166] Then, the server fuses the first recommendation result and the second recommendation result using a fusion sorting algorithm.

[0167] The fusion sorting algorithm will take into account multiple factors, such as the rating of the media source, viewing popularity, and the family's historical viewing habits. Assuming that "XX Daddy" has a high rating on major rating websites and is very popular, it will be given a higher weight in the fusion sorting. Although "Home with XX" has a good rating, it is relatively less popular, so its weight will be slightly lower. Since "XX Panda" is the result of collaborative filtering recommendation, the weight will be determined based on the similarity and viewing frequency of similar users.

[0168] After being processed by the fusion sorting algorithm, the final recommended result for the family user movie list for this family may be: "XX Daddy" ranks first, "XX Panda" ranks second, and "XX at Home" follows closely behind.

[0169] Assuming that it is a weekday, the obtained weekday family portrait shows that this family tends to watch knowledge programs and short news information on weekday evenings.

[0170] The server generates the first recommendation result based on the media vector data and weekday family portraits through the reasoning of the TV media data recommendation model, such as recommending the knowledge lecture "XX Forum" and the news commentary program "News XX".

[0171] Based on the text similarity algorithm, the server recommended the similar program "Kaijiang XX" due to the demand for knowledge-based programs. Through the collaborative filtering algorithm, it was found that other users with similar weekday family portraits often watched the financial news program "Economic XX", which was recommended as the second result.

[0172] When the fusion ranking algorithm is used for fusion, the authority and broad audience base of "XX Forum" are taken into consideration and a higher weight is given. "Opening XX" and "Economic XX" are given corresponding weights according to their characteristics and feedback from similar users.

[0173] The final recommended list may be: "XX Forum" comes first, "Economic XX" comes second, and "Lecture XX" comes last.

[0174] Through such meticulous and comprehensive processing, the server can provide home users with more accurate and diverse TV media recommendation lists that meet their different time needs, thereby improving users' viewing experience and satisfaction.

[0175] In order to more clearly describe the solution provided by the embodiment of the present application, a relatively complete implementation method is provided below. Figure 2 .

[0176] Media asset data vectorization:

[0177] 1.1 Input the local media resource library, including the media resource ID, media resource name, introduction / plot, basic tags, region, release time, etc.;

[0178] 1.2 Preprocess the media data first, and convert the information of each independent field of the media data into natural language description. The media information is first converted into natural language description, so that the context information can be more effectively preserved during vectorization, effectively improving the accuracy of vector retrieval matching.

[0179] 1.3 Use the Chinese word segmentation algorithm to process the media asset description information generated in step 1.2 into Chinese words and generate a word segmentation list;

[0180] 1.4 The FastText model will be used to vectorize media data and write it into the vector library.

[0181] Model training and fine-tuning:

[0182] 2.1 Dataset generation: Generate a model based on the local media library.

[0183] 2.2 Model training: Use llama38B as the pre-trained base model, and use LoRA (Low-rank adaptation) technology + MoE (Mixure of Experts) model training to generate a fine-tuned output model. Compared with training with LoRA alone, using LoRA+MoE for model training can effectively solve the problem of general knowledge forgetting due to fine-tuning of large models, that is, the problem of a sharp decline in model generality, and can effectively maintain the generalization performance of the model while improving the vertical classification capabilities of large models.

[0184] Family portrait generation:

[0185] 3.1 Input the TV user's playback and click behaviors in recent period, including playback / click media ID, media name, time point, playback duration, weekend / holiday mark, etc.;

[0186] 3.2 Preprocess the original family user behavior data to generate a summary description in natural language style; in order to fully consider the differences in family users' viewing behaviors on weekdays and weekends / holidays, the family user behavior description will be differentiated into weekdays and weekends / holidays;

[0187] 3.3 Customize the large model prompt words, input the description of family user behavior, and generate the distribution of TV users' family members and viewing time preferences through model reasoning. Similarly, the prompt word settings will be different for weekdays and weekends / holidays, and finally generate family user portraits for weekdays and holidays;

[0188] Recommendation generation:

[0189] 4.1 Input the user profile data obtained in 3.3 according to different working days, weekends / holidays;

[0190] 4.2 Customize the large model prompt words, combine with the media vector data generated in 1.4, and use the model to infer and generate the recommendation result data for family users throughout the day;

[0191] 4.3 Based on the family user portrait, the traditional portrait preference algorithm based on text similarity and collaborative filtering is used to generate traditional algorithm recommendation results;

[0192] 4.4 uses a custom fusion sorting algorithm to fuse the recommendation results generated by 4.2 and 4.3 to generate the final recommended movie list for home users. Among them, adding the recommendation results of traditional algorithms can effectively solve the problems of narrow recommendation and cold start that may occur in large model recommendations, and increase the diversity and effectiveness of recommendation results.

[0193] Please refer to Figure 3 , Figure 3 A device 110 for recommending and processing television media data provided by an embodiment of the present invention includes:

[0194] The acquisition module 1101 is used to perform vectorization processing on the media asset data according to the local media asset library to generate media asset vector data;

[0195] The training module 1102 is used to train and fine-tune the preset training data set by using the pre-trained base model in combination with the low-rank adaptation strategy and the hybrid expert model to obtain a TV media data recommendation model;

[0196] The recommendation module 1103 is used to obtain the user's TV operation behavior and generate a family portrait based on the user's TV operation behavior; generate recommendation results based on the family portrait and the media vector data through the TV media data recommendation model inference, and use the fusion sorting algorithm to generate the final family user movie list recommendation results.

[0197] It should be noted that the implementation principle of the aforementioned television media data recommendation processing device 110 can refer to the implementation principle of the aforementioned television media data recommendation processing method, which will not be repeated here. It should be understood that the division of the various modules of the above device is only a division of logical functions. In actual implementation, they can be fully or partially integrated into a physical entity, or physically separated. And these modules can all be implemented in the form of software calling through processing elements; they can also be all implemented in the form of hardware; some modules can also be implemented in the form of software calling through processing elements, and some modules can be implemented in the form of hardware. For example, the television media data recommendation processing device 110 can be a separately established processing element, or it can be integrated in a chip of the above device. In addition, it can also be stored in the memory of the above device in the form of program code, and a processing element of the above device can call and execute the function of the above television media data recommendation processing device 110. The implementation of other modules is similar. In addition, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each module above can be completed by an integrated logic circuit of hardware in a processor element or an instruction in the form of software.

[0198] For example, the above modules may be one or more integrated circuits configured to implement the above methods, such as one or more application specific integrated circuits (ASICs), or one or more microprocessors (digital signal processors, DSPs), or one or more field programmable gate arrays (FPGAs), etc. For another example, when a module above is implemented in the form of a processing element scheduling program code, the processing element may be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules may be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0199] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned television media asset data recommendation processing device 110. Figure 4 As shown, Figure 4The computer device 100 provided in the embodiment of the present invention is a structural block diagram. The computer device 100 includes a recommendation processing device 110 for television media asset data, a memory 111 , a processor 112 and a communication unit 113 .

[0200] In order to realize data transmission or interaction, the memory 111, the processor 112 and the communication unit 113 are electrically connected to each other directly or indirectly. For example, the electrical connection between these components can be realized through one or more communication buses or signal lines. The recommendation processing device 110 for television media data includes at least one software function module that can be stored in the memory 111 in the form of software or firmware or fixed in the operating system (OS) of the computer device 100. The processor 112 is used to execute the recommendation processing device 110 for television media data stored in the memory 111, such as the software function modules and computer programs included in the recommendation processing device 110 for television media data.

[0201] An embodiment of the present invention provides a readable storage medium, which includes a computer program. When the computer program is executed, the computer device where the readable storage medium is located is controlled to execute the aforementioned method for recommending processing television media asset data.

[0202] For illustrative purposes, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed. Numerous modifications and variations are possible in accordance with the above teachings. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to suit the intended specific application.

Claims

1. A method for recommending and processing television media data, characterized in that: include: Perform media data vectorization processing based on the local media database to generate media vector data; The pre-trained base model is used to train and fine-tune the preset training data set in combination with the low-rank adaptation strategy and the hybrid expert model to obtain a TV media data recommendation model. Acquire the user's television operation behavior, and generate a family portrait according to the user's television operation behavior; According to the family portrait and the media vector data, the recommendation result is generated by inference through the television media data recommendation model, and the final family user movie list recommendation result is generated by fusion using a fusion sorting algorithm.

2. The method according to claim 1, characterized in that The vectorization processing of media data according to the local media database to generate media vector data includes: Acquire media asset data in the local media asset library, wherein the media asset data includes a media asset ID, a media asset name, a media asset introduction, a media asset plot, basic tags, a region, and a release time; Preprocessing the media asset data to convert independent field information of the media asset data into a natural language description; Perform word segmentation processing on the natural language description to obtain a word segmentation list; The word segmentation list is vectorized using the FastText model to obtain the media asset vector data.

3. The method according to claim 1, characterized in that The method uses the pre-trained base model, combines the low-rank adaptation strategy and the hybrid expert model to train and fine-tune the preset training data set to obtain a TV media data recommendation model, including: Using the llama38B model as the pre-trained base model; The preset training data set is trained and fine-tuned according to the llama38B model combined with the low-rank adaptation strategy and the hybrid expert model to obtain a television media data recommendation model.

4. The method according to claim 1, characterized in that: The acquiring of the user's television operation behavior and generating a family portrait according to the user's television operation behavior includes: Acquire the user's TV operation behavior within a preset time range, the user's TV operation behavior includes user play behavior and user click behavior, and the media asset ID, media asset name, time point, play duration, and weekend and holiday identifier of the media asset data corresponding to the user play behavior and the click behavior; Preprocessing the user's television operation behavior to obtain a user behavior summary description based on a natural language style corresponding to the user's television operation behavior; The family portrait is obtained according to the summary description of the user behavior.

5. The method according to claim 4, characterized in that The user behavior summary description includes a weekday behavior summary description and a weekend and holiday behavior summary description; The step of obtaining the family portrait according to the user behavior summary description includes: The weekday behavior summary description and the weekend and holiday behavior summary description are based on the custom large model prompt words, and the family member distribution and viewing time period preference of the TV user are inferred to obtain the weekday family portrait and the weekend and holiday family portrait.

6. The method according to claim 5, characterized in that The generating of recommendation results by inference of the TV media data recommendation model according to the family portrait and the media vector data, and generating of final family user movie list recommendation results by fusion sorting algorithm, includes: Get current time information; Determine a weekday family portrait or a weekend or holiday family portrait according to the current time information; According to the media vector data and the weekday family portrait or weekend and holiday family portrait, the recommendation result is generated by reasoning through the television media data recommendation model, and the final family user movie list recommendation result is generated by fusion using a fusion sorting algorithm.

7. The method according to claim 6, characterized in that The method of generating a recommendation result by inferring the TV media data recommendation model based on the media vector data and the weekday family portrait or the weekend and holiday family portrait, and generating a final family user movie list recommendation result by fusing the fusion sorting algorithm, includes: Generate a first recommendation result by inferring the television media data recommendation model according to the media asset vector data and the weekday family portrait or the weekend and holiday family portrait; Generate a second recommendation result based on the weekday family portrait or the weekend and holiday family portrait based on text similarity and collaborative filtering algorithms; The first recommendation result and the second recommendation result are fused by using a fusion sorting algorithm to obtain the final home user movie list recommendation result.

8. A device for recommending and processing television media data, characterized in that: include: The acquisition module is used to perform vector processing on the media data according to the local media library to generate media vector data; A training module is used to train and fine-tune a preset training data set using a pre-trained base model in combination with a low-rank adaptation strategy and a hybrid expert model to obtain a TV media data recommendation model; A recommendation module, used to obtain the user's TV operation behavior and generate a family portrait based on the user's TV operation behavior; According to the family portrait and the media vector data, the recommendation result is generated by inference through the television media data recommendation model, and the final family user movie list recommendation result is generated by fusion using a fusion sorting algorithm.

9. A computer device, characterized in that: The computer device comprises a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device executes the method according to any one of claims 1 to 7.

10. A readable storage medium, characterized in that: The readable storage medium includes a computer program, and when the computer program is executed, the computer device where the readable storage medium is located is controlled to execute the method according to any one of claims 1 to 7.

Images