Preference analysis model training method and video recommendation method

Abstract

The embodiment of the invention relates to a preference analysis model training method and a video recommendation method. The preference analysis model training method comprises the steps of obtaining user behavior data of a first user; generating model cue words based on the user behavior data of the first user, the plurality of video feature dimensions and the plurality of preference level features, inputting the model cue words into a large language model, and driving the large language model based on the model cue words to generate at least one group of video preference contents with preference grading tags, the video preference information serves as video preference information of the first user; constructing a training sample based on the user behavior data and the video preference information, inputting the user behavior data in the training sample into a model, outputting a predicted video preference corresponding to the user behavior data by the model, and verifying the predicted video preference by using the video preference information in the training sample, and adjusting parameters of the model according to the verification result until the model reaches a preset convergence condition, and taking the model as a preference analysis model.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and in particular to a training method for a preference analysis model and a video recommendation method. Background Technology

[0002] With the widespread application of large language models in the field of intelligent assistants, users' needs for the content of intelligent assistant responses are becoming increasingly diversified. For example, intelligent assistants for video-based software can recommend videos to users based on their historical viewing behavior data, thereby improving the user experience of video-based software.

[0003] Currently, in order for question-answering models to provide video recommendations that better match user interests when responding, it is necessary to input the user's video preference analysis results into the question-answering model. In related technologies, training samples for the model are usually constructed by manual annotation, and the trained model is used to generate the user's video preference analysis results. Obviously, this manual annotation method not only requires a lot of manpower and time, but also the results of manual annotation are easily affected by the subjective cognition of the annotators, making it difficult to objectively reflect the user's true interest tendencies. Summary of the Invention

[0004] To address, or at least partially address, the aforementioned technical problems, this disclosure provides a method for training a preference analysis model and a video recommendation method.

[0005] This disclosure provides a method for training a preference analysis model, the method comprising: Obtain user behavior data of the first user; wherein, the user behavior data includes historical video playback data and historical interaction behavior data, and the historical interaction behavior data includes the interaction behavior generated by the first user during video playback; Based on the user behavior data of the first user, multiple video feature dimensions, and multiple preference level features, model prompt words are generated. The model prompt words are input into a large language model, and the large language model is driven by the model prompt words to generate at least one set of video preference content with preference level labels, which serves as the video preference information of the first user. The video preference content is used to characterize the video features that the first user is interested in, and the preference level labels of each set of video preference content are used to characterize the degree of interest of the first user in that set of video preference content. Training samples are constructed based on the user behavior data and video preference information of the first user; User behavior data from the training samples is input into the model, which outputs the predicted video preferences corresponding to the user behavior data. The predicted video preferences are verified using the video preference information from the training samples. The parameters of the model are adjusted based on the verification results until the model reaches the preset convergence condition. The model is then used as the preference analysis model.

[0006] In one optional implementation, the step of generating at least one set of video preference content with preference hierarchy labels based on the model prompt words to drive the large language model, as the video preference information of the first user, includes: Based on the model prompt words, the large language model performs clustering processing on the user behavior data according to the multiple video feature dimensions to obtain multiple data sets corresponding to each category; wherein, the data set includes user behavior data under the corresponding category; Based on user behavior data in the dataset corresponding to each category, determine the video preference content corresponding to each category; Based on the video preference content corresponding to the multiple categories, the multiple preference level features, and the user behavior data of the first user, determine the preference level label corresponding to each of the video preference contents; Based on the video preference content and the corresponding preference level tags, the video preference information of the first user is generated.

[0007] In an optional implementation, before generating the video preference information of the first user based on each of the video preference contents and the corresponding preference level tags, the method further includes: Based on the video preference content with preference classification tags and the user behavior data of the first user, the supporting data for the video preference content is generated; The step of generating the video preference information of the first user based on each video preference content and the corresponding preference level tag includes: Based on the video preference content with preference grading tags and the supporting data of the video preference content, the video preference information of the first user is generated.

[0008] In one optional implementation, the step of generating model prompts based on the first user's user behavior data, multiple video feature dimensions, and multiple preference level features includes: Model prompts are generated based on the user behavior data of the first user, multiple preference level features, and reason description standards. Before generating the video preference information of the first user based on the video preference content with the preference classification tags and the supporting data of the video preference content, the method further includes: Based on the video preference content with preference grading tags, the reason description criteria, and the user behavior data of the first user, a preference reason for the video preference content is generated; wherein, the preference reason is used to characterize the reason for determining the video preference content and its corresponding preference grading tags.

[0009] In one optional implementation, the step of driving the large language model based on the model prompt words to cluster the user behavior data according to the multiple video feature dimensions to obtain multiple data sets corresponding to each category includes: Based on the model prompt words, the large language model is driven to perform clustering processing on the user behavior data according to the multiple video feature dimensions to obtain multiple clustering results; wherein, each clustering result includes multiple data sets corresponding to each category; Based on the user behavior data of the first user and the clustering quality requirements, a target clustering result is determined from the multiple clustering results; The step of determining the video preference content corresponding to each category based on user behavior data in the dataset corresponding to each category includes: Based on the user behavior data in the dataset corresponding to each category in the target clustering results, determine the video preference content corresponding to each category.

[0010] This disclosure also provides a video recommendation method, the method comprising: Receive video recommendation query information input by the target user; Acquire user behavior data of the target user; wherein, the user behavior data includes historical video playback data and historical interaction behavior data, and the historical interaction behavior data includes the interactive behavior generated by the target user during video playback; The user behavior data of the target user is input into the preference analysis model, which then determines the target user's video preference information based on the user behavior data. The video preference information includes at least one set of video preference content with preference grading labels. The video preference content characterizes the video features that the target user is interested in, and the preference grading labels for each set of video preference content characterize the target user's degree of interest in that set of video preference content. The preference analysis model is trained using the aforementioned training method for preference analysis models. The video preference information and the video recommendation query information are input into the question answering model. The question answering model determines recommended videos based on the video preference content with preference level tags in each group of video preference information and the video recommendation query information, and outputs the answer content of the video recommendation query information based on the recommended videos.

[0011] This disclosure also provides a training apparatus for a preference analysis model, the apparatus comprising: The first acquisition module is used to acquire user behavior data of the first user; wherein, the user behavior data includes historical video playback data and historical interaction behavior data, and the historical interaction behavior data includes the interaction behavior generated by the first user during video playback; The generation module is used to generate model prompt words based on the user behavior data of the first user, multiple video feature dimensions, and multiple preference level features. The model prompt words are then input into a large language model, which drives the large language model to generate at least one set of video preference content with preference level labels, which serves as the video preference information of the first user. The video preference content is used to characterize the video features that the first user is interested in, and the preference level labels of each set of video preference content are used to characterize the degree of interest of the first user in that set of video preference content. The construction module is used to construct training samples based on the user behavior data and video preference information of the first user; The training module is used to input user behavior data from the training samples into the model, output the predicted video preferences corresponding to the user behavior data from the model, verify the predicted video preferences using the video preference information from the training samples, adjust the parameters of the model according to the verification results, and use the model as a preference analysis model when the model reaches the preset convergence condition.

[0012] This disclosure also provides a video recommendation device, the device comprising: The receiving module is used to receive video recommendation query information input by the target user; The second acquisition module is used to acquire user behavior data of the target user; wherein, the user behavior data includes historical video playback data and historical interaction behavior data, and the historical interaction behavior data includes the interactive behavior generated by the target user during video playback; The first determining module is used to input the user behavior data of the target user into the preference analysis model, and the preference analysis model determines the video preference information of the target user based on the user behavior data; wherein, the video preference information includes at least one set of video preference content with preference level labels, the video preference content is used to characterize the video features of the target user's interest, the preference level label of each set of video preference content is used to characterize the degree of interest of the target user for that set of video preference content, and the preference analysis model is trained by the above-mentioned preference analysis model training method; The second determining module is used to input the video preference information and the video recommendation query information into the question answering model. The question answering model determines the recommended video based on the video recommendation query information and the video preference information for each group of video preference content with preference level tags, and outputs the answer content of the video recommendation query information based on the recommended video.

[0013] This disclosure also provides an electronic device, which includes: Processor and memory; The processor executes the method described in any embodiment of this disclosure by calling programs or instructions stored in the memory.

[0014] This disclosure also provides a computer-readable storage medium that stores a program or instructions that cause a computer to perform the methods described in any embodiment of this disclosure.

[0015] The training method for the preference analysis model provided in this disclosure combines user behavior data, multiple video feature dimensions, and multiple preference level features to generate model prompt words, driving a large language model to generate video preference content with preference level labels. This serves as the first user's video preference information. Since this video preference information can not only accurately represent the characteristics of videos the user is interested in, but also quantify the degree of user interest through level labels, the preference analysis model trained using this video preference information can recommend videos that better match the user's preferences. Compared to manual annotation, automatically generating video preference information using a large language model reduces sample construction time and generates video preference information that better reflects the user's true interests.

[0016] Furthermore, the video recommendation method provided in this embodiment inputs the user behavior data of the target user into a preference analysis model. The preference analysis model determines the video preference information of the target video based on the user behavior data. This video preference information includes at least one set of video preference content with preference level tags. This allows the question-answering model to determine recommended videos based on the video recommendation query information and each set of video preference content with preference level tags in the video preference information. The model then outputs the answer content for the video query information based on the recommended videos. Compared to directly recommending videos to the target user based on user behavior data, this embodiment can recommend videos that better match the target user's preferences based on video preference information, thereby improving the user experience of video software. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0018] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a training method for a preference analysis model provided in this embodiment of the present disclosure; Figure 2 A flowchart illustrating a video recommendation method provided in an embodiment of this disclosure; Figure 3 A flowchart for generating video preference information and intensity levels is provided as an embodiment of this disclosure; Figure 4 A schematic diagram of the structure of a training device for a preference analysis model provided in an embodiment of this disclosure; Figure 5 This is a schematic diagram of the structure of a video recommendation device provided in an embodiment of the present disclosure; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0020] To better understand the above-mentioned objectives, features, and advantages of this disclosure, the solutions disclosed herein will be described in further detail below. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0021] Numerous specific details are set forth in the following description in order to provide a full understanding of this disclosure, but this disclosure may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some, and not all, of the embodiments of this disclosure.

[0022] With the continuous development of data processing technology, video software is becoming increasingly feature-rich. It can recommend video files to users based on their past viewing history. For example, the intelligent assistant in video software can recommend videos that match a user's interests based on their historical viewing behavior data, thereby improving the user experience.

[0023] Currently, in order for question-answering models to provide video recommendations that better match user interests when responding, it is necessary to input the user's video preference analysis results into the question-answering model. In related technologies, training samples for the model are usually constructed by manual annotation, and the trained model is used to generate the user's video preference analysis results. However, this manual annotation method not only requires a lot of manpower and time, but the results of manual annotation are also easily affected by the subjective cognition of the annotators, making it difficult to objectively reflect the user's true interest tendencies.

[0024] Based on this, the present disclosure provides a training method for a preference analysis model. The method includes: firstly, acquiring user behavior data of a first user, including historical playback data and historical interaction behavior data; then, generating model prompt words based on the first user's user behavior data, multiple video feature dimensions, and multiple preference level features, and inputting the model prompt words into a large language model; driving the large language model based on the model prompt words to generate at least one set of video preference content with preference level labels, as the first user's video preference information; next, constructing training samples based on the first user's user behavior data and video preference information, and inputting the user behavior data in the training samples into the model; the model outputting the predicted video preferences corresponding to the user behavior data; verifying the predicted video preferences using the video preference information in the training samples; adjusting the model parameters according to the verification results; and finally, when the model reaches a preset convergence condition, using the model as a preference analysis model.

[0025] In this embodiment, by combining user behavior data, multiple video feature dimensions, and multiple preference level features to generate model prompt words, a large language model is driven to generate video preference content with preference level labels. This video preference information serves as the first user's video preference information. This video preference information can not only accurately represent the characteristics of videos that the user is interested in, but also quantify the degree of user interest through level labels. Therefore, the method of automatically generating video preference information using a large language model provided in this embodiment, compared to manual annotation, can reduce sample construction time and generate video preference information that better reflects the user's true interests.

[0026] The preference analysis model training method provided in this disclosure can be applied to question-and-answer scenarios of intelligent assistants provided by video software. This preference analysis model training method can be executed by a preference analysis model training device, which can be implemented in software and / or hardware. This device can be integrated into electronic devices with plot information extraction capabilities, such as mobile phones, PDAs, tablets, laptops, or desktop computers.

[0027] Figure 1 A flowchart illustrating a training method for a preference analysis model provided in this embodiment of the disclosure. See also... Figure 1The training method specifically includes: S110. Obtain the user behavior data of the first user.

[0028] The user behavior data includes historical video playback data and historical interaction behavior data, and the historical interaction behavior data includes the interaction behavior generated by the first user during video playback.

[0029] Historical video playback data includes information about videos watched by the first user within a preset historical time period. For example, basic information about long or short videos watched by the first user in the past year. Historical video playback data can include basic video information such as video name, type, year of release, and playback progress. In addition, for long videos watched by the first user within the preset historical time period, the corresponding video information also includes plot summary, plot highlights, main actors, rating, and other information.

[0030] Historical interaction behavior data includes the interactive behaviors generated by the first user during video playback, which can be used to reflect the first user's video preferences for the currently playing video. Historical interaction behavior data may include, for example, the first user's likes, favorites, comments, and reposts during video playback.

[0031] Historical interaction behavior data can also include search keywords entered by the first user within a preset historical time period, such as video-related search keywords entered by the first user in the search box in the past three months. Each piece of historical interaction behavior data can include search keywords, clicked content, and search time.

[0032] S120. Based on the user behavior data of the first user, multiple video feature dimensions and multiple preference level features, generate model prompt words, input the model prompt words into the large language model, and drive the large language model to generate at least one set of video preference content with preference level labels based on the model prompt words, as the video preference information of the first user.

[0033] The video preference content is used to characterize the video features that the first user is interested in, and the preference level label for each group of video preference content is used to characterize the degree of interest of the first user in that group of video preference content.

[0034] The video feature dimensions are determined based on the user behavior data of the first user. The video feature dimensions may include features such as region, production company, source, year, story background, subject matter elements, actors, director, target audience, core character image, story theme, content style, art style, tone, atmosphere, viewing experience, quality, content freshness, and dissemination.

[0035] Preference level features are used to characterize the first user's interest in video features of different dimensions. The higher the intensity level, the higher the first user's interest in the video corresponding to that feature dimension. The lower the intensity level, the lower the target user's interest in the video corresponding to that feature dimension. The intensity levels from strong to weak can be: strong preference, moderate preference, and latent preference.

[0036] In this embodiment of the disclosure, after obtaining the user behavior data of the first user, model prompt words are first generated based on the user behavior data of the first user, multiple video feature dimensions, and multiple preference level features. Then, the model prompt words are input into the large language model, and the large language model is driven by the model prompt words to generate at least one set of video preference content with preference level labels, which serves as the video preference information of the first user.

[0037] Among them, video preference features are used to characterize the video features that the first user is interested in, such as "classic nonsensical comedy movies starring A". This video preference feature includes multiple video features, such as actor A, movie genre, comedy type, etc.

[0038] Preference rating labels for video content are used to characterize the degree of interest of the first user in the video content. Preference rating labels may include strong preference, moderate preference, and potential preference.

[0039] In practical applications, the degree of preference for video playback can be divided into multiple preference levels by combining multiple dimensions such as viewing progress information, search frequency information, number of episodes played, number of viewed content, and bullet comments.

[0040] The viewing progress information represents the target user's progress in watching the video. For example, a viewing progress of 300% for a particular video means the target user has watched the video three times. The search frequency information represents the number of times the target user has searched for and clicked on videos. For example, the target user may have searched for "youth" twice and watched two videos with the theme of "youth" within the past three days.

[0041] For example, a strong preference refers to video preference information where the viewing progress information matches a preset first progress range (e.g., two or more videos with a playback progress greater than 300%), or video preference information where the search frequency information matches a preset first frequency range (e.g., searched twice in the last three days). A moderate preference refers to video preference information where the viewing progress information matches a second progress range (e.g., one video with a playback progress greater than 100%, or three or more videos with a playback progress between 50% and 100%). A potential preference refers to video preference information where only historical bullet comment behavior data or search frequency information matches a second frequency range.

[0042] As can be seen, the embodiments of this disclosure output at least one set of video preference content with preference level labels through a preference analysis model, so as to use preference level labels to characterize the degree of interest of the first user in the video features indicated by the video preference content.

[0043] S130. Construct training samples based on the user behavior data and video preference information of the first user.

[0044] In this embodiment of the disclosure, after generating video preference information using a large language model, training samples are constructed based on the user behavior data and video preference information of the first user. That is, the training samples include user behavior data and video preference information with corresponding relationships.

[0045] S140. Input the user behavior data in the training samples into the model, and have the model output the predicted video preference corresponding to the user behavior data. Validate the predicted video preference using the video preference information in the training samples. Adjust the parameters of the model according to the validation results until the model reaches the preset convergence condition. Then, use the model as the preference analysis model.

[0046] The predicted video preferences are the inferences about user video preferences calculated by the model based on input user behavior data and its own algorithm. In the early stages of training, the predicted video preferences may deviate significantly from the actual preferences. The preset convergence criteria are the quantitative standards for determining whether the model training is complete. They are pre-set goals; meeting the preset convergence criteria indicates that the model's predictive ability has stabilized and met the target, and further parameter tuning is unnecessary.

[0047] In this embodiment of the disclosure, training samples containing user behavior data and corresponding real video preference information are input into the model to be trained so that the model can output the predicted video preference corresponding to the user behavior data. Then, the real video preference information in the training samples is used to compare the predicted video preference of the model to determine the accuracy of the predicted video preference. Based on the verification results after comparison, the internal parameters of the model are adjusted to reduce the prediction bias.

[0048] Repeat the above steps of input, prediction, verification, and parameter tuning until the model's prediction results meet the preset convergence conditions, such as the prediction accuracy reaching the target or the error stabilizing within the preset threshold. At this point, the model can be used as the final usable preference analysis model.

[0049] The training method for the preference analysis model provided in this embodiment first acquires user behavior data of a first user, including historical playback data and historical interaction behavior data. Then, based on the user behavior data of the first user, multiple video feature dimensions, and multiple preference level features, model prompt words are generated and input into a large language model. The large language model is driven by the model prompt words to generate at least one set of video preference content with preference level labels, which serves as the video preference information of the first user. Next, training samples are constructed based on the user behavior data and video preference information of the first user, and the user behavior data in the training samples is input into the model. The model outputs the predicted video preferences corresponding to the user behavior data. The predicted video preferences are verified using the video preference information in the training samples, and the model parameters are adjusted according to the verification results until the model reaches the preset convergence condition. The model is then used as the preference analysis model.

[0050] As can be seen, since the preference analysis model in this embodiment is trained based on user behavior data and video preference information samples with corresponding relationships, it can be used to mine video preference information from user behavior data that better reflects the types of videos that the first user is interested in.

[0051] In some embodiments, based on the above technical solutions, the specific steps for generating the first user's video preference information based on the model prompt word-driven large language model may include: First, a large language model driven by model prompt words is used to cluster user behavior data according to multiple video feature dimensions, resulting in multiple datasets corresponding to each category. Then, based on the user behavior data in each category's dataset, the video preference content corresponding to each category is determined. Next, based on the video preference content corresponding to each category, multiple preference level features, and the user behavior data of the first user, preference level labels corresponding to each video preference content are determined. Finally, based on each video preference content and its corresponding preference level label, the video preference information of the first user is generated.

[0052] The clustered dataset includes user behavior data for the corresponding categories.

[0053] This disclosure does not impose specific limitations on the model and training method. The model may include machine learning models, large language models, etc., and the training method may include low-rank adaptive supervised fine-tuning (LoRA SFT) method, etc. During the training process, the model may be trained in multiple rounds until a preset number of rounds is reached or convergence occurs.

[0054] The model to be trained may include a causal language model architecture using a decoder, specifically consisting of an input embedding layer, positional encoding, a Transformer decoding layer (core layer, 64 layers in total), and an output layer.

[0055] In the process of clustering user behavior data, the video features of the user behavior data are first extracted based on the pre-configured video attributes. Then, user behavior data samples with common video features are clustered into one category, resulting in multiple data sets corresponding to each category. Based on the user behavior data in the data set corresponding to each category, the video preference content corresponding to each category is determined.

[0056] In practical applications, for a dataset corresponding to a certain category, multiple user behavior data points must all satisfy the video features of the video preference content corresponding to that category. Furthermore, any user behavior data sample in a dataset cannot appear in other datasets.

[0057] For example, user behavior data can be clustered based on the following model prompt words that drive a large language model: "Below is the user behavior data of the first user. Please perform feature analysis and summarize the video content preferences of the first user."

[0058] {Historical video playback data}; {Historical interaction behavior data}; [Concept Explanation]: Video feature dimensions are summarized based on information provided in user behavior data and your knowledge, including but not limited to the following aspects: 1. Region: American dramas, British dramas, Japanese dramas, Eastern dramas, etc.; 2. Source: Adapted from classic works, adapted from true events, etc.; 3. Era: Modern, Contemporary, Modern, Ancient; 4. Story background; 5. Thematic elements 6. Actors: Idol, talent, good looks, etc. 7. Director and screenwriter: Famous director and famous screenwriter; 8. Target audience: Children, under 0 years old, 0-3 years old, 4-6 years old, female-oriented, male-oriented, family-friendly, niche audience, etc.

[0059] [Summary Requirements]: 1. Extract the core features of each piece of content from user behavior data.

[0060] 2. Group content with common characteristics into one category, prioritizing the grouping of two or more dramas with the same characteristics into one category.

[0061] 3. When classifying, features can only be extracted from the following aspects: whether the highlights of multiple dramas have commonalities, whether the core story reflected in the plot of multiple dramas has commonalities, and whether multiple dramas are characterized as high quality, new and popular, newly released, old, or classic, reflecting users' preferences for content quality and whether the content is new or old.

[0062] 4. Describe the video features in one sentence. Note: Features should be detailed and must be analyzed from two or more dimensions; do not use a single type as a clustering feature.

[0063] In some embodiments, to further improve the accuracy of clustering results, a higher value can be set for the divergence of the large language model before clustering user behavior data based on model prompt words to encourage the large language model to output data sets corresponding to multiple categories.

[0064] In other embodiments, after the large language model driven by model prompt words performs clustering processing on user behavior data according to multiple video feature dimensions to obtain data sets corresponding to multiple categories, negative sampling can be used to improve the accuracy of clustering results.

[0065] Specifically, based on the user behavior data of the first user and the clustering quality requirements, the target clustering result is determined from multiple clustering results; then, based on the user behavior data in the dataset corresponding to each category in the target clustering result, the video preference content corresponding to each category is determined.

[0066] The target clustering result is the clustering result that meets the clustering quality requirements among multiple clustering results, such as a high degree of matching between the clustered categories and the corresponding user behavior data. For example, when determining the target result from multiple clustering results, the corresponding prompt word can be designed as follows: "{User behavior data of the first user};" {Multiple clustering results}; Task Objective: Please evaluate the multiple clustering results provided above based on the user behavior data of the first user, and select the one with the highest quality. The evaluation content includes, but is not limited to: Preference features should be accurately expressed and should only be used to summarize content features; user behavior and preference levels should not be mixed in. The core content features of user movie-watching behavior under each preference feature are highly consistent with the preference feature, and there is no incorrect clustering; No viewing content with a completion rate of over 50% should be missed; Preference expressions cannot contain "or" relationships; they must be broken down or modified based on user behavior. For example: "Family dramas with a frontier / urban background." To refine the features, cross-analysis must be conducted from two or more dimensions, for example: "intergenerational collision travel reality show"; "Using a single type as a clustering feature is prohibited."

[0067] For example, video preference information that meets the following conditions can be determined as the final video preference information used for model training: there are no similar preference features in the video preference information, for example, the video preference information cannot contain both "suspenseful reasoning modern TV dramas" and "suspenseful crime-solving TV dramas" at the same time; and there is no identical user historical behavior data in the clustering result set corresponding to any two video preference information.

[0068] In other embodiments, before generating the first user's video preference information based on the video preference content and the corresponding preference rating tags, the process may further include: First, based on the video preference content with preference grading labels and the user behavior data of the first user, supporting data for the video preference content is generated; then, based on the video preference content with preference grading labels and the supporting data for the video preference content, the video preference information of the first user is generated.

[0069] The supporting data includes user behavior description data that supports the corresponding video preference content, which is used to characterize the historical behavior data of the first user for the videos indicated by the video interest features in the video preference content.

[0070] In other embodiments, after generating the video preference information of the first user based on the video preference content with preference level labels and the supporting data of the video preference content, the preference reasons of the video preference content can also be generated based on the user behavior data of the first user, multiple preference level features and reason description standards.

[0071] Specifically, firstly, model prompt words are generated based on the user behavior data of the first user, multiple preference level features, and reason description standards. Then, based on the model prompt words, the large language model is driven to generate preference reasons for the video preference content according to the video preference content with preference level labels, reason description standards, and the user behavior data of the first user.

[0072] Among them, the reason description criteria are used to describe the criteria for determining video preference content, as well as the criteria for determining the preference level labels corresponding to the video preference content.

[0073] The rationale for video preference content is used to characterize the reasons for determining the video preference content and its corresponding preference level label. For example, the recommendation reason for the video preference content "classic slapstick comedy movie starring A" could be "You have watched 7 comedy videos starring A in the past three months, so we recommend video A to you."

[0074] In this embodiment of the disclosure, the video preference information of the first user output by the large language model driven by model prompt words may include video preference content, preference level tags corresponding to the video preference content, supporting data of the video preference content, and the reasons for the preference content.

[0075] The video preference information of the first user is shown in the following example: Video preference content: "Classic nonsensical comedy movies starring Xiao A"; Preference rating label: "Strong preference"; Reason for preference: "Watched 11 videos starring Xiao A, and the first user actively searched for movies related to Xiao A 7 times, meeting the criteria of [more than 2 movies with high completion rate and supplementary searches with the same characteristics]"; Supporting data: "Title: Kung Fu, Genre: Movie, Playback Completion Rate: 939%; Title: The Great Swordsman, Genre: Movie, Playback Completion Rate: 656%".

[0076] In practical applications, if historical viewing behavior data is directly input into the question-and-answer model of an intelligent assistant, and the model is used to output recommended videos, the accuracy of the recommended videos will be low due to the limited processing power of the question-and-answer model, thus affecting the user's experience with video software.

[0077] Therefore, embodiments of this disclosure provide a video recommendation method. Figure 2 This is a flowchart of a video recommendation method provided in an embodiment of this disclosure. See also... Figure 2 The video recommendation method specifically includes: S210: Receive video recommendation query information input by the target user.

[0078] The video recommendation query information refers to the natural language text input by the target user to obtain recommended videos, such as the text content "Recommend a few videos for me," etc. After receiving the video recommendation query information input by the target user, the intelligent assistant in the video software will generate corresponding answer content based on the video recommendation query information. The target user can be any user.

[0079] S220. Obtain user behavior data of the target user.

[0080] User behavior data includes historical video playback data and historical interaction behavior data. Historical interaction behavior data refers to the interactive behaviors generated by the target user during video playback.

[0081] Historical video playback data includes information about videos watched by target users within a preset historical time period. For example, basic information about long or short videos watched by target users in the past year can be included. Historical video playback data can include basic video information such as video name, type, release year, and playback progress. In addition, for long videos watched by target users within a preset historical time period, the corresponding video information also includes plot summary, plot highlights, main actors, rating, and other information.

[0082] Historical interaction behavior data includes at least one of the following: historical bullet screen behavior data and historical search behavior data.

[0083] Historical bullet screen behavior data includes the target user's evaluation information on multiple consecutive frames of video. Each piece of historical bullet screen behavior data can include information such as bullet screen content, movie review name and posting date. Bullet screen content can include text, images, emoticons and other types of content.

[0084] Historical search behavior data includes search keywords entered by target users within a preset historical time period, such as video-related search keywords entered by target users in the search box in the past three months. Each piece of historical search behavior data can include search keywords, clicked content, and search time.

[0085] As can be seen, after receiving video recommendation query information input by the user, this embodiment of the present disclosure obtains the user behavior data of the target user so that the target user can be further analyzed and processed to determine the types of videos that the target user is interested in.

[0086] S130. Input the target user's user behavior data into the preference analysis model, and the preference analysis model determines the target user's video preference information based on the user behavior data.

[0087] The video preference information includes at least one set of video preference content with preference sub-labels, which is used to characterize the video features that the target user is interested in. For example, if the video preference information is "classic slapstick comedy movies starring Xiao A", it means that the target user is more interested in comedy videos and videos starring actor Xiao A.

[0088] The preference rating labels for each set of video preferences are used to characterize the target user's level of interest in that set of video preferences. The preference analysis model is trained using the same training method described above.

[0089] The preference analysis model can be trained based on pre-acquired training samples, which include user behavior data and video preference information of the first user with corresponding relationships.

[0090] In this embodiment of the disclosure, since the preference analysis model is trained based on user behavior data and video preference information with corresponding relationships, and is used to mine the types of videos that users are interested in from user behavior data, by inputting the user behavior data of the target user into the preference analysis model, and after processing by the trained preference analysis model, video preference information that better reflects the types of videos that the target user is interested in can be extracted.

[0091] S240. Input the video preference information and video recommendation query information into the question answering model. The question answering model determines the recommended video based on the video preference content of each group of videos with preference level labels in the video recommendation query information and video preference information, and outputs the answer content of the video recommendation query information based on the recommended video.

[0092] In this embodiment of the disclosure, after determining the target user's video preference information using a preference analysis model, the video preference information and video recommendation query information are input into a question-answering model. The question-answering model then determines recommended videos and outputs the answer content for the video recommendation query information based on the recommended videos.

[0093] Since the data input to the question-answering model in this embodiment is the video preference information of the target user determined by the user behavior data of the target user using the preference analysis model, it can characterize the types of videos that the target user is interested in. Therefore, compared with the method of directly recommending videos to the target user based on user behavior data, this embodiment can use video preference information to recommend recommended videos that are more in line with the target user's preferences, thereby improving the accuracy of recommended videos.

[0094] like Figure 3 The diagram shown is a flowchart of a method for generating video preference information and intensity levels according to an embodiment of this disclosure.

[0095] After acquiring the user behavior data of the first user, the system first performs clustering on the user behavior data according to multiple video feature dimensions based on the model prompt word-driven large language model, resulting in multiple data sets corresponding to each category. Then, based on the user behavior data in the data set corresponding to each category, the system determines the video preference content corresponding to each category. Furthermore, based on the video preference content corresponding to multiple categories, multiple preference level features, and the user behavior data of the first user, the system determines the preference classification labels corresponding to each video preference content. Additionally, based on the video preference content with preference classification labels and the user behavior data of the first user, the system generates supporting data for the video preference content. Finally, based on the video preference content with preference classification labels, the reason description representation, and the user behavior data of the first user, the system generates preference reasons for the video preference content. Finally, the system generates the video preference information of the first user based on each video preference content and its corresponding preference classification labels, supporting data, and preference reasons.

[0096] Next, training samples are constructed based on the user behavior data and video preference information of the first user, and the model is trained based on the training samples to obtain the preference analysis model.

[0097] In the subsequent application of the model, upon receiving video recommendation query information input by the target user, the model first acquires the target user's user behavior data. This data is then input into the aforementioned preference analysis model, which determines the target user's video preference information based on the user behavior data. The video preference information includes at least one set of video preference content with preference grading labels. This video preference content characterizes the video features that the target user is interested in, and the preference grading labels for each set of video preference content characterize the target user's degree of interest in that set of video preference content.

[0098] Next, the video preference information and video recommendation query information are input into the question answering model. The question answering model determines the recommended video based on the video preference content with preference level labels in each group of video recommendation query information and video preference information, and outputs the answer content of the video recommendation query information based on the recommended video.

[0099] As can be seen, by inputting video preference content with preference classification tags into the question answering model, the question answering model can recommend recommended videos that users are more interested in based on the video preference content and its corresponding preference classification tags, thereby further improving the user experience.

[0100] Figure 4 This is a schematic diagram of the structure of a training device for a preference analysis model provided in an embodiment of this disclosure. Figure 4 As shown, the training device 400 includes: The first acquisition module 401 is used to acquire user behavior data of the first user; wherein, the user behavior data includes historical video playback data and historical interaction behavior data, and the historical interaction behavior data includes the interaction behavior generated by the first user during video playback; The generation module 402 is used to generate model prompt words based on the user behavior data of the first user, multiple video feature dimensions, and multiple preference level features, input the model prompt words into a large language model, and drive the large language model to generate at least one set of video preference content with preference level labels based on the model prompt words, as the video preference information of the first user; wherein, the video preference content is used to characterize the video features that the first user is interested in, and the preference level labels of each set of video preference content are used to characterize the degree of interest of the first user in that set of video preference content; Construction module 403 is used to construct training samples based on the user behavior data and video preference information of the first user; The training module 404 is used to input user behavior data from the training samples into the model, output the predicted video preferences corresponding to the user behavior data from the model, verify the predicted video preferences using the video preference information from the training samples, adjust the parameters of the model according to the verification results, and use the model as a preference analysis model when the model reaches the preset convergence condition.

[0101] In the training device for a preference analysis model provided in this embodiment, model prompts are generated by combining user behavior data, multiple video feature dimensions, and multiple preference level features. This drives a large language model to generate video preference content with preference level labels, serving as the video preference information for the first user. Since this video preference information can not only accurately represent the characteristics of videos that the user is interested in, but also quantify the degree of user interest through level labels, the preference analysis model trained using this video preference information can recommend videos that better match the user's preferences. Compared to manual annotation, automatically generating video preference information using a large language model reduces sample construction time and generates video preference information that better reflects the user's true interests.

[0102] In some embodiments, the generation module includes: The clustering processing submodule is used to drive the large language model based on the model prompt words to perform clustering processing on the user behavior data according to the multiple video feature dimensions, so as to obtain multiple data sets corresponding to each category; wherein, the data set includes user behavior data under the corresponding category; The determination submodule is used to determine the video preference content corresponding to each category based on user behavior data in the dataset corresponding to each category. Based on the video preference content corresponding to the multiple categories, the multiple preference level features, and the user behavior data of the first user, determine the preference level label corresponding to each of the video preference contents; The first generation submodule is used to generate the video preference information of the first user based on the video preference content and the corresponding preference level tags.

[0103] In some embodiments, the generation module further includes: The second generation submodule is used to generate supporting data for the video preference content based on the video preference content with preference level labels and the user behavior data of the first user. The first generation submodule is specifically used for: Based on the video preference content with preference grading tags and the supporting data of the video preference content, the video preference information of the first user is generated.

[0104] In some embodiments, the generation module includes: The third generation submodule is used to generate model prompt words based on the user behavior data of the first user, multiple preference level features, and reason description standards. The third generation submodule is used to generate a preference reason for the video preference content based on the video preference content with preference level tags, the reason description standard, and the user behavior data of the first user; wherein the preference reason is used to characterize the determination reason of the video preference content and its corresponding preference level tags.

[0105] In some embodiments, the clustering processing submodule is specifically used for: Based on the model prompt words, the large language model is driven to perform clustering processing on the user behavior data according to the multiple video feature dimensions to obtain multiple clustering results; wherein, each clustering result includes multiple data sets corresponding to each category; Based on the user behavior data of the first user and the clustering quality requirements, a target clustering result is determined from the multiple clustering results; The determining submodule is specifically used for: Based on the user behavior data in the dataset corresponding to each category in the target clustering results, determine the video preference content corresponding to each category.

[0106] Figure 5 This is a schematic diagram of the structure of a video recommendation device provided in an embodiment of this disclosure. Figure 5 As shown, the video recommendation device 500 includes: The receiving module 501 is used to receive video recommendation query information input by the target user; The second acquisition module 502 is used to acquire user behavior data of the target user; wherein, the user behavior data includes historical video playback data and historical interaction behavior data, and the historical interaction behavior data includes the interactive behavior generated by the target user during video playback; The first determining module 503 is used to input the user behavior data of the target user into the preference analysis model, and the preference analysis model determines the video preference information of the target user based on the user behavior data; wherein, the video preference information includes at least one set of video preference content with preference level labels, the video preference content is used to characterize the video features of interest of the target user, the preference level label of each set of video preference content is used to characterize the degree of interest of the target user for that set of video preference content, and the preference analysis model is trained by the above-mentioned preference analysis model training method; The second determining module 504 is used to input the video preference information and the video recommendation query information into the question answering model, and the question answering model determines the recommended video based on the video recommendation query information and the video preference information for each group of video preference content with preference level tags, and outputs the answer content of the video recommendation query information based on the recommended video.

[0107] In a video recommendation device provided in this disclosure, the historical behavior data of a target user is input into a preference analysis model. The preference analysis model extracts video preference information that reflects the types of videos the target user is interested in. Based on this video preference information and video recommendation query information, the question-answering model can determine recommended videos that better match the target user's preferences. Compared to directly recommending videos to the target user based on user behavior data, this disclosure can recommend recommended videos that better match the target user's preferences based on video preference information, thereby improving the user's experience with video software.

[0108] The video recommendation device provided in this disclosure can execute the video recommendation method provided in any embodiment of this disclosure, and has the corresponding functional modules and beneficial effects of executing the method.

[0109] It is worth noting that in the embodiments of the video recommendation device described above, the various sub-modules are divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional sub-module are only for easy differentiation and are not used to limit the scope of protection of this disclosure.

[0110] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Figure 6 As shown, the electronic device 600 includes one or more processors 601 and memory 602.

[0111] The processor 601 may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 600 to perform desired functions.

[0112] The memory 602 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 601 may execute the program instructions to implement the video recommendation method of the embodiments of this disclosure described above and / or other desired functions. Various content, such as historical behavior data corresponding to target users, may also be stored in the computer-readable storage medium.

[0113] In one example, electronic device 600 may further include an input device 603 and an output device 604, these components being interconnected via a bus system and / or other forms of connection mechanisms (not shown). The input device 603 may include, for example, a keyboard, a mouse, etc. The output device 604 may output various information to the outside, including identified recommended videos, reasons for recommendation, etc. The output device 604 may include, for example, a display, speakers, a printer, and a communication network and its connected remote output devices, etc.

[0114] Of course, for the sake of simplicity, Figure 6 Only some of the components of the electronic device 600 relevant to this disclosure are shown, omitting components such as buses, input / output interfaces, etc. In addition, the electronic device 600 may include any other suitable components depending on the specific application.

[0115] In addition to the methods and devices described above, embodiments of this disclosure may also be computer program products, including computer program instructions that, when executed by a processor, cause the processor to perform the video recommendation method provided in the embodiments of this disclosure.

[0116] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this disclosure. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on a user's computing device, partially on a user's computing device, as a standalone software package, partially on a user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0117] Furthermore, embodiments of this disclosure may also be computer-readable storage media storing computer program instructions that, when executed by a processor, cause the processor to perform the video recommendation method provided in embodiments of this disclosure.

[0118] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0119] It should be noted that the terminology used in this disclosure is for the purpose of describing specific embodiments only and is not intended to limit the scope of this disclosure. As shown in this specification and claims, unless the context clearly indicates otherwise, words such as "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. The term "and / or" includes any one and all combinations of one or more of the associated listed items. Relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another entity or operation and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, or apparatus. Without further limitation, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element.

[0120] The above description is merely a specific embodiment of this disclosure, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A training method for a preference analysis model, characterized in that, include: Obtain user behavior data of the first user; wherein, the user behavior data includes historical video playback data and historical interaction behavior data, and the historical interaction behavior data includes the interaction behavior generated by the first user during video playback; Based on the user behavior data of the first user, multiple video feature dimensions, and multiple preference level features, model prompt words are generated. The model prompt words are input into a large language model, and the large language model is driven by the model prompt words to generate at least one set of video preference content with preference level labels, which serves as the video preference information of the first user. The video preference content is used to characterize the video features that the first user is interested in, and the preference level labels of each set of video preference content are used to characterize the degree of interest of the first user in that set of video preference content. Training samples are constructed based on the user behavior data and video preference information of the first user; User behavior data from the training samples is input into the model, which outputs the predicted video preferences corresponding to the user behavior data. The predicted video preferences are verified using the video preference information from the training samples. The parameters of the model are adjusted based on the verification results until the model reaches the preset convergence condition. The model is then used as the preference analysis model.

2. The method according to claim 1, characterized in that, The process of generating at least one set of video preference content with preference hierarchy labels based on the model prompt words, as the video preference information of the first user, includes: Based on the model prompt words, the large language model performs clustering processing on the user behavior data according to the multiple video feature dimensions to obtain multiple data sets corresponding to each category; wherein, the data set includes user behavior data under the corresponding category; Based on user behavior data in the dataset corresponding to each category, determine the video preference content corresponding to each category; Based on the video preference content corresponding to the multiple categories, the multiple preference level features, and the user behavior data of the first user, determine the preference level label corresponding to each of the video preference contents; Based on the video preference content and the corresponding preference level tags, the video preference information of the first user is generated.

3. The method according to claim 2, characterized in that, Before generating the video preference information of the first user based on each video preference content and the corresponding preference level tag, the method further includes: Based on the video preference content with preference classification tags and the user behavior data of the first user, the supporting data for the video preference content is generated; The step of generating the video preference information of the first user based on each video preference content and the corresponding preference level tag includes: Based on the video preference content with preference grading tags and the supporting data of the video preference content, the video preference information of the first user is generated.

4. The method according to claim 3, characterized in that, The method of generating model prompts based on the user behavior data of the first user, multiple video feature dimensions, and multiple preference level features includes: Model prompts are generated based on the user behavior data of the first user, multiple preference level features, and reason description standards. Before generating the video preference information of the first user based on the video preference content with the preference classification tags and the supporting data of the video preference content, the method further includes: Based on the video preference content with preference grading tags, the reason description criteria, and the user behavior data of the first user, a preference reason for the video preference content is generated; wherein, the preference reason is used to characterize the reason for determining the video preference content and its corresponding preference grading tags.

5. The method according to claim 2, characterized in that, The method, based on the model prompt words, drives the large language model to cluster the user behavior data according to multiple video feature dimensions, obtaining multiple data sets corresponding to each category, including: Based on the model prompt words, the large language model is driven to perform clustering processing on the user behavior data according to the multiple video feature dimensions to obtain multiple clustering results; wherein, each clustering result includes multiple data sets corresponding to each category; Based on the user behavior data of the first user and the clustering quality requirements, a target clustering result is determined from the multiple clustering results; The step of determining the video preference content corresponding to each category based on user behavior data in the dataset corresponding to each category includes: Based on the user behavior data in the dataset corresponding to each category in the target clustering results, determine the video preference content corresponding to each category.

6. A video recommendation method, characterized in that, include: Receive video recommendation query information input by the target user; Acquire user behavior data of the target user; wherein, the user behavior data includes historical video playback data and historical interaction behavior data, and the historical interaction behavior data includes the interactive behavior generated by the target user during video playback; The user behavior data of the target user is input into the preference analysis model, and the preference analysis model determines the video preference information of the target user based on the user behavior data; wherein, the video preference information includes at least one set of video preference content with preference level labels, the video preference content is used to characterize the video features of interest of the target user, the preference level label of each set of video preference content is used to characterize the degree of interest of the target user for that set of video preference content, and the preference analysis model is trained by the preference analysis model training method of any one of claims 1-5; The video preference information and the video recommendation query information are input into the question answering model. The question answering model determines recommended videos based on the video preference content with preference level tags in each group of video preference information and the video recommendation query information, and outputs the answer content of the video recommendation query information based on the recommended videos.

7. A training device for a preference analysis model, characterized in that, include: The first acquisition module is used to acquire user behavior data of the first user; wherein, the user behavior data includes historical video playback data and historical interaction behavior data, and the historical interaction behavior data includes the interaction behavior generated by the first user during video playback; The generation module is used to generate model prompt words based on the user behavior data of the first user, multiple video feature dimensions, and multiple preference level features. The model prompt words are then input into a large language model, which drives the large language model to generate at least one set of video preference content with preference level labels, which serves as the video preference information of the first user. The video preference content is used to characterize the video features that the first user is interested in, and the preference level labels of each set of video preference content are used to characterize the degree of interest of the first user in that set of video preference content. The construction module is used to construct training samples based on the user behavior data and video preference information of the first user; The training module is used to input user behavior data from the training samples into the model, output the predicted video preferences corresponding to the user behavior data from the model, verify the predicted video preferences using the video preference information from the training samples, adjust the parameters of the model according to the verification results, and use the model as a preference analysis model when the model reaches the preset convergence condition.

8. A video recommendation device, characterized in that, include: The receiving module is used to receive video recommendation query information input by the target user; The second acquisition module is used to acquire user behavior data of the target user; wherein, the user behavior data includes historical video playback data and historical interaction behavior data, and the historical interaction behavior data includes the interactive behavior generated by the target user during video playback; The first determining module is used to input the user behavior data of the target user into the preference analysis model, and the preference analysis model determines the video preference information of the target user based on the user behavior data; wherein, the video preference information includes at least one set of video preference content with preference level labels, the video preference content is used to characterize the video features of interest of the target user, the preference level label of each set of video preference content is used to characterize the degree of interest of the target user for that set of video preference content, and the preference analysis model is trained by the training method of the preference analysis model according to any one of claims 1-5; The second determining module is used to input the video preference information and the video recommendation query information into the question answering model. The question answering model determines the recommended video based on the video recommendation query information and the video preference information for each group of video preference content with preference level tags, and outputs the answer content of the video recommendation query information based on the recommended video.

9. An electronic device, characterized in that, The electronic device includes: Processor and memory; The processor executes the training method of the preference analysis model as described in any one of claims 1 to 5 or the video recommendation method as described in claim 6 by calling the program or instructions stored in the memory.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program or instructions that cause a computer to perform a training method for a preference analysis model as described in any one of claims 1 to 5 or a video recommendation method as described in claim 6.

11. A computer program product, characterized in that, The computer program product is used to implement the training method of the preference analysis model according to any one of claims 1-6 or the video recommendation method according to claim 6.

Images