Short video recommendation method and system based on large language model

Abstract

The invention provides a short video recommendation method and system based on a large language model, a storage medium and electronic equipment, and relates to the field of machine learning. According to the method, video content is integrated to generate a large language model, an interest description large language model and a circle breaking strategy description large language model, and a recommendation system architecture for assisting a user in circle breaking is designed. The method comprises the following steps: firstly, accurately identifying multi-dimensional features of a video and user interests and preferences, and dividing different interesting communities; secondly, executing a video recommendation strategy of a double-track system, and respectively obtaining short video recommendation lists based on interests and auxiliary circle breaking; and then, fusing the two short video recommendation lists, deploying the proportion of the broken circle short videos in recommendation according to the broken circle willingness of the target user and the decision score of the manager, generating a final short video recommendation list, and exploring the short video content in which the short videos are potentially interested. Furthermore, a large language model is generated based on the title, the title of the broken-circle short video is designed, and the click desire and watching interest of the target user are stimulated.

Description

Technical Field

[0001] The present invention relates to the field of machine learning, and in particular to a short video recommendation method, system, storage medium and electronic device based on a large language model. Background Art

[0002] In today's digital age, short videos, with their real-time interactivity, wide user participation and instant content dissemination, have become a core medium for the public to express themselves, enjoy entertainment and obtain diverse information.

[0003] With the rapid development of the short video industry and the continuous expansion of the user scale, the problem of information overload is becoming increasingly serious. The massive amount of short video content and diverse user needs make it a key issue to filter out information that meets user interests. In related technologies, the existing personalized recommendation algorithms have solved this problem to a certain extent and improved the user experience, but the risk of information homogeneity has also emerged. When users enjoy the convenience brought by personalized recommendations, they are often easily trapped in a solidified information circle, which limits the breadth and depth of information acquisition.

[0004] Therefore, how to balance the relationship between personalized recommendations and diversified information has become an important topic that requires in-depth thinking and exploration. Summary of the invention

[0005] 1. Technical issues to be resolved

[0006] In view of the deficiencies in the prior art, the present invention provides a short video recommendation method, system, storage medium and electronic device based on a large language model, which solves the technical problem of how to balance the relationship between personalized recommendations and diversified information.

[0007] (II) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0009] A short video recommendation method based on a large language model, comprising:

[0010] Obtain a short video set and a user set containing the target user;

[0011] A pre-trained BERT-based short video content generation large language model is used to generate a description text for each short video, and the TF-IDF algorithm is used to extract keywords from the description text to obtain several tags for each short video, and the tag set of the short video set is obtained after optimization and merging;

[0012] Count the historical short video sets of interest to each user within a preset historical time period, combine the tag set to obtain several interest tags of each user, and use the K-means algorithm for clustering to determine the interest community to which each user belongs;

[0013] Compare the similarity between the target user's interest tag and the tag of any short video to generate a first short video recommendation list based on the user's personal interest; in the interest community to which the target user belongs, adopt a collaborative filtering method based on neighboring users to generate a second short video recommendation list based on the interest community; adopt a Boosting algorithm to fuse the first short video recommendation list and the second short video recommendation list to obtain a third short video recommendation list based on interest;

[0014] Locate the communities closest to and farthest from the interest community to which the target user belongs, and construct corresponding candidate short video recommendation lists; merge the candidate short video recommendation lists of the nearest and farthest communities, and allocate the number of recommended short videos based on the distance to obtain a fourth short video recommendation list that assists in breaking the circle;

[0015] Based on the target user's willingness to break the circle and the manager's decision score, the proportion of the fourth short video recommendation list is calculated, and combined with the third short video recommendation list to obtain the final short video recommendation list; wherein the pre-trained interest description large language model based on the Transformer deep learning architecture and the breaking circle strategy description large language model are used, combined with the evaluation results given by the manager, to obtain the manager's decision score.

[0016] Preferably, a pre-trained large language model for title generation based on a Transformer deep learning architecture is used to generate the title of each short video in the short video recommendation list that originates from the fourth short video recommendation list.

[0017] Preferably, the method uses a pre-trained BERT-based short video content to generate a large language model to generate a description text for each short video, and uses a TF-IDF algorithm to extract keywords from the description text to obtain a number of tags for each short video, and optimizes and merges them to obtain a tag set for the short video set; including:

[0018] For any short video, a motion detection algorithm is used to decode the continuous video stream into a series of key frames F o , where o represents the index of the key frame;

[0019] Calculate each key frame F o The color histogram defines the color space and color channels, and counts the number of pixels of each color value in each color channel to form the color histogram feature vector v color,o ;

[0020] The gray-level co-occurrence matrix algorithm is used to extract the texture feature vector v by statistically analyzing the joint probability distribution of pixel pairs in the gray-level image at the preset direction and distance. texture,o ;

[0021] Use the Canny edge detection algorithm to detect the edges in the image, and combine the Hough transform to extract the shape features to obtain the shape feature set S o , and converted into shape feature vector v shape,o ;

[0022] Extract the audio features of the short video and convert the dubbing in the short video into a text sequence T audio ;

[0023] Fusion visual feature vector set {v color,o ,v texture,o ,v shape,o} and the converted text sequence T audio , use the pre-trained BERT-based short video content to generate a large language model for encoding, generate a hidden state vector, and generate a descriptive text D through a fully connected layer or classifier;

[0024] Preprocessing the descriptive text D, extracting keywords from the preprocessed descriptive text D using a TF-IDF algorithm, and using each extracted keyword as a candidate tag;

[0025] Based on the semantic relationship between the candidate tags, a machine learning algorithm is used to optimize the tags to obtain several tags {l1,l2,…,l m}, where m represents the total number of tags of the short video;

[0026] Merge the labels of all short videos to obtain the label set L of the short video set total .

[0027] Preferably, the method comprises: counting the historical short video sets of interest of each user within a preset historical time period, combining the tag set, obtaining a number of interest tags of each user, and clustering using the K-means algorithm to determine the interest community to which each user belongs; including:

[0028] The user set is defined as U = {u1,u2,…,u n}, where n represents the total number of users; define user u i The collection of historically interesting short videos within the preset historical time period is where i m Represents user u i The total number of historical short videos of interest; for each short video v j ∈V i , define the tag set of the short video as L total A tag set representing the short video set;

[0029] For user u i , build the corresponding label count dictionary

[0030]

[0031] in, Indicates that label l is in user u i History of Interesting Short Video Collection V i The number of times it appears in the equation; I represents the indicator function. 1 when it is, otherwise 0;

[0032] Select the tag whose occurrence frequency is greater than the preset threshold as the user u i Merge all users’ interest tags and remove duplicates to obtain the user interest tag set L top ;

[0033] Clustering using K-means algorithm:

[0034]

[0035] Among them, SSE represents the sum of squares of distances from all users to their corresponding cluster centers; K represents the set number of clusters; C c represents the set of users in the cth cluster; μ c is the center of the cth cluster;

[0036] Taking minimizing SSE as the optimization goal, iteratively update the cluster center and reallocate users to the nearest cluster center to determine user u i Interest group C pi .

[0037] Preferably, the comparing the similarity between the interest tag of the target user and the tag of any short video to generate a first short video recommendation list based on the user's personal interests comprises:

[0038] For target users Define a feature vector If and only if L top When {k} is the user's interest tag, Equal to 1; for short video v j , define its label as If and only if L top When {k} is the tag of the video, is equal to 1; where |L to[ | represents the number of tags in the user’s interest tag set, Ltop {k} represents the kth tag in the user's interest tag set;

[0039] The following formula is used to compare the similarity between the target user's interest tag and the tag of any short video:

[0040]

[0041] Where sim represents the similarity function, · represents the dot product, and ||·|| represents the Euclidean length of the vector;

[0042] By comparing the similarities, we find a collection of videos that are similar to the target user’s interests and sort them according to the similarity scores to generate a first short video recommendation list V based on the user’s personal interests. CBR .

[0043] Preferably, in the interest community to which the target user belongs, a collaborative filtering method based on neighboring users is adopted to generate a second short video recommendation list based on the interest community; comprising:

[0044] Calculate the similarity between users and determine the target users The neighbor user set of

[0045] Utilize the short videos v that each neighbor user in the interest community to which the target user belongs has watched or rated p The target users are predicted by weighted average method. Potential interest in unrated short videos; calculated as follows:

[0046]

[0047] in, Indicates the target user For short video v p The predicted score of the target user is N(u), N(u) represents the interest community to which the target user belongs, r u′_vp Represents the neighbor user u′’s response to the short video v p The actual rating of

[0048] The short videos are sorted based on the predicted scores, and several short videos with the highest scores are selected as the second short video recommendation list V based on the interest community. UCF .

[0049] Preferably, the process of obtaining the target user's willingness to break the circle includes:

[0050] For target users Interest groups Sort the interesting tags of the cluster center, define the first several tags with the most occurrences as the mainstream tags of the community; define Function, calculate the target user The proportion of videos not belonging to the mainstream tag of the community watched during the preset historical period is used as a measure of the target users. An indicator of personal breaking circle activity;

[0051] Combined with the Jaccard coefficient, define function, as a measure of target users Interest groups Indicators of the tendency of social groups to break out of their circles;

[0052] Among them, the target users Interest groups The number of mainstream tags is used as the base number; interest communities will be counted in two adjacent preset historical time periods. The change of tags of watching short videos is taken as the change number; The function is the ratio of the difference between the base number and the variation to the sum of the base number and the variation;

[0053] Define the target user The breaking circle willingness score function:

[0054]

[0055] in, is a value between 0 and 1; α, β represent the corresponding weight coefficients respectively, and α+β=1, α, β∈(0,1).

[0056] Preferably, the process of obtaining the manager decision score includes:

[0057] Using the tags of the third short video recommendation list as input to the interest description large language model to generate a text description of the community interest; and using the tags of the fourth short video recommendation list as input to the breaking circle strategy description large language model to generate a text description of the breaking circle strategy;

[0058] Based on the text description of the community interests and the text description of the breaking circle strategy, combined with the evaluation results given by the manager, the manager's decision score is defined as S decide , the value is between 0 and 1.

[0059] Preferably, based on the target user's willingness to break the circle and the manager's decision score, the proportion of the fourth short video recommendation list is calculated, and combined with the third short video recommendation list to obtain the final short video recommendation list; including:

[0060] Calculate the short video recommendation list V Recommend The number of short videos is:

[0061] NUM(V last )=γ1NUM(V UCF )+γ2NUM(V breakout )

[0062]

[0063] Among them, γ1, γ2 represent the corresponding weight coefficients, and γ1+γ2=1, γ1, γ2∈(0,1); NUM(V UCF ),NUM(V breakout ) represent the third short video recommendation list V UCF And the fourth short video recommendation list V breakout The number of short videos;

[0064] Select V according to the ratio UCF and V breakout The short videos with the highest to lowest ratings are combined to obtain the final recommendation list V Recommend .

[0065] A short video recommendation system based on a large language model, comprising:

[0066] An acquisition module, used to acquire a short video set and a user set including a target user;

[0067] The recognition module is used to generate a description text of each short video using a pre-trained BERT-based short video content generation large language model, and extract keywords from the description text using a TF-IDF algorithm to obtain a number of tags for each short video, and obtain a tag set of the short video set after optimization and merging;

[0068] and for counting the historical short video sets of interest of each user within a preset historical time period, combining the tag set to obtain a number of interest tags of each user, and clustering using the K-means algorithm to determine the interest community to which each user belongs;

[0069] The recommendation module is used to compare the similarity between the interest tag of the target user and the tag of any short video, and generate a first short video recommendation list based on the user's personal interest; in the interest community to which the target user belongs, a collaborative filtering method based on neighboring users is used to generate a second short video recommendation list based on the interest community; and a Boosting algorithm is used to fuse the first short video recommendation list and the second short video recommendation list to obtain a third short video recommendation list based on interest;

[0070] and for respectively locating the communities closest to and farthest from the interest community to which the target user belongs, and constructing corresponding candidate short video recommendation lists; integrating the candidate short video recommendation lists of the nearest and farthest communities, and allocating the number of recommended short videos based on the distance, so as to obtain a fourth short video recommendation list that assists in breaking the circle;

[0071] The optimization module is used to calculate the proportion of the fourth short video recommendation list based on the target user's willingness to break the circle and the manager's decision score, and obtain the final short video recommendation list in combination with the third short video recommendation list; wherein the pre-trained interest description large language model based on the Transformer deep learning architecture and the breaking circle strategy description large language model are used, combined with the evaluation results given by the manager, to obtain the manager's decision score.

[0072] A storage medium stores a computer program for short video recommendation based on a large language model, wherein the computer program enables a computer to execute the short video recommendation method as described above.

[0073] An electronic device, comprising:

[0074] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include a program for executing the short video recommendation method as described above.

[0075] (III) Beneficial effects

[0076] The present invention provides a short video recommendation method, system, storage medium and electronic device based on a large language model. Compared with the prior art, the present invention has the following beneficial effects:

[0077] The present invention integrates the video content to generate a large language model, an interest description large language model, and a breaking circle strategy description large language model, and designs a recommendation system architecture that assists users in breaking the circle. First, accurately identify the multi-dimensional characteristics of the video and the user's interest preferences, and divide the users into different interest communities; secondly, implement a dual-track video recommendation strategy to obtain interest-based and auxiliary breaking circle short video recommendation lists respectively; then, merge the two short video recommendation lists, and adjust the proportion of breaking circle short videos in the recommendation according to the target user's breaking circle willingness and the manager's decision score, generate the final short video recommendation list, and discover potential short video content of interest. While realizing personalized recommendations, the present invention provides high-quality and diversified short video recommendations for target users, which helps them break out of the information cocoon and get in touch with a wider world. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0079] Figure 1 An architecture diagram of a recommendation system for assisting users to break out of their circles provided by an embodiment of the present invention;

[0080] Figure 2 A flowchart of a short video recommendation method based on a large language model provided by an embodiment of the present invention;

[0081] Figure 3 A flowchart of another short video recommendation method based on a large language model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0082] 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 are clearly and completely described. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0083] The embodiments of the present application solve the technical problem of how to balance the relationship between personalized recommendations and diversified information by providing a short video recommendation method, system, storage medium and electronic device based on a large language model.

[0084] The technical solution in the embodiment of the present application is to solve the above technical problems, and the overall idea is as follows:

[0085] Large Language Model (LLM) is a model based on the Transformer architecture

[48] , which is deeply trained on massive text data and built with a large parameter set. In recent years, with the improvement of computing power and the continuous innovation of algorithms, research on large language models has made significant progress worldwide. Well-known foreign models such as ChatGPT and LLaMA, as well as domestic models such as Wenxin Yiyan ERNIE bot and PanGu, have demonstrated strong performance in practical applications.

[0086] In this context, the embodiment of the present invention uses the large language model technology to design a recommendation system architecture that helps users break through the circle (break through the original solidified information circle), such as Figure 1As shown, it is hoped that through technical means, users can explore diversified information; specifically:

[0087] Step 1: Interest Identification

[0088] In interest recognition, the first task is to accurately identify the multi-dimensional features of the video and the user's interest preferences. The visual features of the video are extracted and the text information is converted into audio features to achieve comprehensive feature integration. Subsequently, a large language model based on BERT is introduced to generate short video content. With its powerful multimodal data processing capabilities, a variety of visual elements and audio texts are deeply analyzed and integrated to generate detailed video description texts, providing solid text support for the subsequent formulation of video tags. Through TF-IDF technology and machine learning algorithm optimization, an efficient and deduplicated video tag set is constructed to support diverse video application needs.

[0089] Based on video feature recognition and user interest analysis, we further build user portraits. Through data analysis and clustering technology, we divide users into different interest groups and extract unique interest tags for each group. This process not only helps us to deeply understand the user's interest preferences, but also provides strong support for more personalized video recommendations and content distribution.

[0090] Step 2: Video recommendation

[0091] Based on video tags and user characteristics, a dual-track video recommendation strategy is designed. This strategy is closely centered around the in-depth analysis of video tags and user characteristics. On the one hand, in order to enhance the accuracy of recommendations, personalized recommendations based on interests are made. Based on the user's existing interests, through content-based and community-based collaborative filtering algorithms, short video content closely related to the user's current interests is accurately identified and pushed. On the other hand, in order to stimulate users' out-of-circle behavior, out-of-circle short video recommendations are made.

[0092] Based on the interest community to which the user belongs, we explore the interests of their neighboring and distant communities, build a candidate set of short videos that break the circle, and distribute videos according to the distance ratio between communities, so as to recommend to users those non-traditional video content that may arouse their interest although they do not belong to their original interest circle. At the same time, based on the video recommendation results, we train large language models for interest description and breaking circle strategy description. The former can accurately portray the personalized interests of users and communities, while the latter explains the selection logic and characteristics of breaking circle short videos, providing a scientific basis for subsequent managers' decision-making.

[0093] Step 3: Recommended Optimization

[0094] The two recommendation lists are combined, and the proportion of breaking circle short videos in the recommendation is adjusted according to the characteristics of user community breaking circle behavior, individual breaking circle willingness and manager's strategic considerations. User breaking circle willingness is based on historical behavior and community dynamics, and manager decision-making ensures that the strategy is compliant and efficient. Using the trained interest description big language model and the breaking circle strategy description big language model, a detailed description of the hybrid strategy is generated to assist managers in making decisions that conform to market and user needs. Finally, by weighted mixing of high-scoring videos, a personalized and diverse recommendation list is created.

[0095] Furthermore, in the generation stage of the recommendation list, the embodiment of the present invention also gives full play to the creative potential and generation ability of the large language model. Combining the personalized interest preferences of users with the unique charm of the video content, a large language model for title generation is designed to tailor attractive titles for the short videos in the recommendation list that break the circle. These titles capture the user's focus of interest and emotional resonance to stimulate the user's desire to click and interest in watching, thereby improving the attractiveness and conversion rate of the recommended content, attracting users to interact with the recommended videos, and helping users to break the circle.

[0096] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0097] Embodiment 1:

[0098] like Figure 2 As shown, an embodiment of the present invention provides a short video recommendation method based on a large language model, comprising:

[0099] S1. Obtain a short video set and a user set including the target user;

[0100] S2. Use the pre-trained BERT-based short video content generation large language model to generate a description text for each short video, and use the TF-IDF algorithm to extract keywords from the description text to obtain several tags for each short video, and optimize and merge to obtain a tag set for the short video set;

[0101] S3, counting the historical short video sets of interest of each user in a preset historical time period, combining the tag set to obtain a number of interest tags of each user, and clustering using the K-means algorithm to determine the interest community to which each user belongs;

[0102] S4, comparing the similarity between the interest tag of the target user and the tag of any short video, generating a first short video recommendation list based on the user's personal interest; in the interest community to which the target user belongs, using a collaborative filtering method based on neighboring users, generating a second short video recommendation list based on the interest community; using a Boosting algorithm to fuse the first short video recommendation list and the second short video recommendation list, obtaining a third short video recommendation list based on interest;

[0103] S5. Locate the communities closest to and farthest from the interest community to which the target user belongs, respectively, and construct corresponding candidate short video recommendation lists; merge the candidate short video recommendation lists of the nearest and farthest communities, and allocate the number of recommended short videos based on the distance to obtain a fourth short video recommendation list that assists in breaking the circle;

[0104] S6. Based on the target user's willingness to break the circle and the manager's decision score, calculate the proportion of the fourth short video recommendation list, and combine it with the third short video recommendation list to obtain the final short video recommendation list; wherein the pre-trained interest description large language model based on the Transformer deep learning architecture and the breaking circle strategy description large language model are used, combined with the evaluation results given by the manager to obtain the manager's decision score.

[0105] The embodiments of the present invention integrate video content to generate a big language model, an interest description big language model, and a breaking circle strategy description big language model, and design a recommendation system architecture to assist users in breaking the circle. While realizing personalized recommendations, it provides high-quality and diversified short video recommendations for target users, helping them to break out of the information cocoon and access a wider world.

[0106] In an optional embodiment, if Figure 3 As shown, the embodiment of the present invention also includes:

[0107] S7. Use a pre-trained large language model for title generation based on the Transformer deep learning architecture to generate the title of each short video in the short video recommendation list that originates from the fourth short video recommendation list.

[0108] The embodiment of the present invention further integrates a large language model for title generation, and tailors attractive titles specifically for the short videos that break the circle in the recommendation list, so as to stimulate users' desire to click and interest in watching, thereby improving the attractiveness and conversion rate of the recommended content, attracting users to interact with the recommended videos, and helping users to break the circle.

[0109] Next, we will combine Figure 1 The following are the steps of the above solution:

[0110] In step S1, a short video set and a user set including a target user are obtained.

[0111] The embodiment of the present invention focuses on the short video platform, and the default platform covers a wide range of user groups, including individuals from different regions, ages, and professional backgrounds. In addition, while watching short videos on the platform, users can perform interactive operations such as likes and ratings.

[0112] This step obtains a short video set consisting of multiple short videos from the platform, and collects short videos containing target users. The user set U={u1,u2,…,u n}, where n represents the total number of users.

[0113] In step S2, a pre-trained BERT-based short video content generation large language model is used to generate a description text for each short video, and the TF-IDF algorithm is used to extract keywords from the description text to obtain several tags for each short video, which are then optimized and merged to obtain a tag set for the short video set.

[0114] It should be noted that compared with other media forms, the uniqueness of video is that it has both dynamic image and sound features. Extracting visual features in a video means extracting the features of each key frame of the video.

[0115] Accordingly, this step performs video feature recognition in the interest recognition phase, specifically including:

[0116] For any short video, a motion detection algorithm is used to decode the continuous video stream into a series of key frames F o , where o represents the index of the key frame. Calculate each key frame F o The color histogram defines the color space and color channels, and counts the number of pixels of each color value in each color channel to form the color histogram feature vector v color,o The gray-level co-occurrence matrix algorithm is used to extract the texture feature vector v by statistically analyzing the joint probability distribution of pixel pairs in the gray-level image at the preset direction and distance. texture,o The Canny edge detection algorithm is used to detect the edges in the image, and the shape features are extracted by combining the Hough transform to obtain the shape feature set S o , and converted into shape feature vector v shape,o . Extract the audio features of the short video and convert the dubbing in the short video into a text sequence T audio .

[0117] Based on feature extraction, the visual feature vector set {v color,o ,v texture,o ,v shape,o} and the converted text sequence T audio, using the pre-trained BERT-based short video content to generate a large language model for encoding, generating a hidden state vector, and generating a descriptive text D through a fully connected layer or classifier. The descriptive text D can contain key information such as the theme, key events, emotional tendencies, and styles of the video. Among them, the BERT model captures the contextual relationship in the input sequence through the self-attention mechanism and generates a hidden state vector.

[0118] Preprocessing the descriptive text D includes steps such as word segmentation, stop word removal, stem extraction or word form restoration. Next, the TF-IDF algorithm is used to extract keywords from the preprocessed descriptive text D, and each extracted keyword is used as a candidate tag. The processing method is as follows:

[0119]

[0120] Where TF(k,D) is the frequency of word t in D, and k represents the index of word t; ∑ k TF(t,D) is the total frequency of all words in D (normalization factor); log represents the logarithmic function; |D total | is the total number of all descriptive texts D; {D′∈D total : t∈D′}| is the number of descriptive texts D containing word t.

[0121] Considering the semantic relationship between tags, avoiding the generation of semantically similar or repeated tags, using machine learning algorithms such as support vector machines to optimize tags, and obtaining several tags {l1,l2,…,l m}, where m represents the total number of tags of the short video.

[0122] Merge the labels of all short videos to obtain the label set L of the short video set total ,This tag set provides rich metadata support for applications such as video retrieval, recommendation systems, and content classification.

[0123] In step S3, the historical short video sets of interest to each user within a preset historical time period are counted, and several interest tags of each user are obtained in combination with the tag set, and clustering is performed using the K-means algorithm to determine the interest community to which each user belongs.

[0124] This step performs the interest community division in the interest identification phase, that is, based on the video tag set generated in the previous step, the user is divided into communities based on their viewing interests; specifically, it includes:

[0125] First, define user u i The collection of historically interesting short videos within the preset historical time period is For example, define user u iThe collection of videos that were watched, or "liked," within two weeks is where i m Represents user u i The total number of short videos with historical interest / likes; for each short video v j ∈V i , define the tag set of the short video as L total A tag set representing the short video set.

[0126] Then extract the user interest tags. i , build the corresponding label count dictionary

[0127]

[0128] in, Indicates that label l is in user u i History of Interesting Short Video Collection V i The number of times it appears in the equation; I represents the indicator function. 1 if the value is 0, otherwise it is 0.

[0129] In order to reduce the impact of the video on the user, only the impact of the tag on the user is considered, and the tag whose number of occurrences is greater than the preset threshold is selected as the user u i tags of interest, such as tags that appear more than 5 times.

[0130] According to the user's interest tags, all users in the user set U are divided into communities. Merge all users' interest tags and remove duplicates to obtain the user interest tag set L top .

[0131] Clustering using K-means algorithm:

[0132]

[0133] Among them, SSE represents the sum of squares of distances from all users to their corresponding cluster centers; K represents the set number of clusters; C c represents the set of users in the cth cluster; μ c is the center of the cth cluster;

[0134] Taking minimizing SSE as the optimization goal, iteratively update the cluster center and reallocate users to the nearest cluster center to determine user u i Interest group C pi .

[0135] Finally, the clustering results can be output, all users can be assigned to K clusters, and user u can be defined as i The community is Cpi .

[0136] Furthermore, the labels of the cluster centers can be sorted, for example, the five labels with the highest number of occurrences can be defined as community C. p The mainstream label is (l p1 ,l p2 ,l p3 ,l p4 ,l p5 ), where l p1 ,l p2 ,l p3 ,l p4 ,l p5 ∈L total .

[0137] Based on the precise depiction of user portraits and the clear division of the user's community, the embodiment of the present invention designs a dual-track video recommendation strategy. This strategy not only focuses on the core interests of the target users, but also screens short video content that suits their preferences through algorithms to ensure the accuracy and satisfaction of the recommendations; at the same time, it explores the interest boundaries of the target users and selects those short videos that may arouse the curiosity of the target users and guide them to cross the existing interest categories. Such a recommendation mechanism aims to promote the diversified development of the target users' interests, encourage the target users to try new things, thereby broadening their horizons and enjoying a more colorful video viewing experience. For a detailed introduction, please refer to steps S4 to S5:

[0138] In step S4, the similarity between the interest tag of the target user and the tag of any short video is compared to generate a first short video recommendation list based on the user's personal interest; in the interest community to which the target user belongs, a collaborative filtering method based on neighboring users is used to generate a second short video recommendation list based on the interest community; the Boosting algorithm is used to fuse the first short video recommendation list and the second short video recommendation list to obtain a third short video recommendation list based on interest.

[0139] The embodiment of the present invention recommends videos to users based on their interests. The content-based recommendation algorithm is a commonly used recommendation system algorithm that recommends other short videos that are similar to the interests of the target user by analyzing the content features of the short videos that the target user liked in the past. This algorithm mainly relies on the attributes of the target user itself, without considering the behavior of other users. In this scenario, the similarity between the target user's personal interest tags and the video tags is compared to recommend videos that better match their interests to users.

[0140] Accordingly, this step performs interest-based recommendation in the video recommendation phase, specifically including:

[0141] For target users Define a feature vector If and only if L top When {k} is the user's interest tag, Equal to 1; for short video v j , define its label as If and only if L top When {k} is the tag of the video, is equal to 1; where |L top | represents the number of tags in the user’s interest tag set, L top {k} represents the kth tag in the user's interest tag set.

[0142] The following formula is used to compare the similarity between the target user's interest tag and the tag of any short video:

[0143]

[0144] Among them, sim represents the similarity function, · represents the dot product, and ||·|| represents the Euclidean length of the vector.

[0145] By comparing the similarities, we find a collection of videos that are similar to the target user’s interests and sort them according to the similarity scores to generate a first short video recommendation list V based on the user’s personal interests. CBR .

[0146] Furthermore, in the context of interest-based community division, the interest similarity of the target user within the interest community to which he belongs is higher than the similarity between different groups. In order to achieve personalized recommendation, a collaborative filtering method based on neighbor users is adopted, including:

[0147] Calculate the similarity between users and determine the target users The neighbor user set of

[0148] Utilize the short videos v that each neighbor user in the interest community to which the target user belongs has watched or rated p The target users are predicted by weighted average method. Potential interest in unrated short videos; calculated as follows:

[0149]

[0150] in, Indicates the target user For short video v p The predicted score of the target user is N(u), N(u) represents the interest community to which the target user belongs, r u′ _ vp Represents the neighbor user u′’s response to the short video v p actual rating.

[0151] The short videos are sorted based on the predicted scores, and several short videos with the highest scores are selected as the second short video recommendation list V based on the interest community. UCF .

[0152] Finally, the Boosting integrated learning idea is used to integrate V CBR and V UCF , the two recommendation lists are regarded as the outputs of two "weak recommenders". According to the results of performance evaluation, a weight is assigned to each weak recommender. The weight is directly assigned according to the size of the performance index. The better the performance, the greater the weight. The prediction results of these weak recommenders are weighted and combined with simple linear weighting to finally obtain the interest-based recommendation list V normal .

[0153] In step S5, the communities closest to and farthest from the interest community to which the target user belongs are located respectively, and corresponding candidate short video recommendation lists are constructed; the candidate short video recommendation lists of the nearest and farthest communities are merged, and the number of recommended short videos is allocated based on the distance to obtain a fourth short video recommendation list that assists in breaking the circle.

[0154] For target users Make full use of their interest communities in personalized recommendations The characteristics are designed to not only satisfy their existing interests, but also explore their potential interest boundaries, that is, to achieve "breaking the circle".

[0155] Accordingly, this step performs the recommendation of breaking circle short videos in the video recommendation link. Specifically, a two-step strategy is adopted to construct a breaking circle short video set, including:

[0156] First, identify the cluster centers and Recent communities The distance between the two centers is d min This community's video viewing interests are Close, but different, most likely to generate target users Haven't watched the video but liked it.

[0157] Second, in order to ensure the target users You can see the most comprehensive types of videos, explore the possibility of breaking the circle, and further locate the social group The farthest community The distance between the two centers is d max , it and The distance between them is the farthest, and their interests are the most different, representing the target users. New areas that may not have been touched before can help break the circle.

[0158] for and Construct candidate short video recommendation lists V respectively min and V max The two candidate short video recommendation lists are the videos with the highest to lowest ratings in the two communities.

[0159] Based on this, we build the fourth short video recommendation list V to help break the circle. breakout , the number of videos in the two candidate video sets in the new breaking circle short video set satisfies:

[0160] NUM(V breakout )=NUM(V min )×r+NUM(V max )×s (6)

[0161] Among them, r and s are proportional coefficients, r is the number of times the community will recommend to the target user. The proportion of videos in the recommendation list is s, and the farthest community will recommend it to the target user. The proportion of videos in the recommendation list is r+s=1, r,s∈(0,1). This design aims to ensure that the recommendation list retains the content familiar to the target users while also introducing enough novel elements to stimulate the target users' curiosity and desire to explore.

[0162] Considering that in k-means clustering, the distance between communities usually reflects the degree of difference in their interests or preferences. A closer distance means that the interests between communities are more similar, while a farther distance indicates a greater difference in interests. Therefore, allocating the number of recommended videos according to the distance can make the recommendation list introduce novelty and diversity while maintaining a certain similarity. Therefore, let r:s = d max :d min , It can directly reflect the similarities and differences between communities, making the recommendation strategy easier to understand and implement. According to the obtained ratio, the two selected lists are merged to obtain the fourth short video recommendation list V breakout .

[0163] In step S6, based on the target user's willingness to break the circle and the manager's decision score, the proportion of the fourth short video recommendation list is calculated, and combined with the third short video recommendation list to obtain the final short video recommendation list; wherein the pre-trained interest description large language model based on the Transformer deep learning architecture and the breaking circle strategy description large language model are used, combined with the evaluation results given by the manager to obtain the manager's decision score.

[0164] This step performs the recommendation optimization phase to mix the normal recommendation list V constructed based on user interests. normalList of videos that will help users break the circle V breakout ; Specifically include:

[0165] First, obtain the target user's willingness to break out of the circle; including:

[0166] For target users Interest groups Sort the interesting tags of the cluster center, define the first several tags with the most occurrences as the mainstream tags of the community; define Function, calculate the target user The proportion of videos not belonging to the mainstream tag of the community watched during the preset historical period is used as a measure of the target users. An indicator of personal breaking circle activity.

[0167] Combined with the Jaccard coefficient, define function, as a measure of target users Interest groups An indicator of the tendency of a community to break out of its circle.

[0168] Among them, the target users Interest groups The number of mainstream tags is used as the base number; interest communities will be counted in two adjacent preset historical time periods. The change of tags of watching short videos is taken as the change number; The function is the ratio of the difference between the base and the variation to the sum of the base and the variation.

[0169] In fact, here The function is essentially a coefficient of change, which uses no change as a benchmark to calculate the similarity between each type of change and no change. The higher the similarity, the smaller the change, and the less willing the community members are to try and accept diverse content; the lower the similarity, the greater the change, indicating that the community members are more willing to try and accept diverse content.

[0170] The target user is defined by integrating the target user’s personal circle-breaking activity and community circle-breaking tendency through weighted summation. The score function of willingness to break the circle:

[0171]

[0172] in, is a value between 0 and 1; α, β represent the corresponding weight coefficients respectively, and α+β=1, α, β∈(0,1).

[0173] It can be understood that the embodiments of the present invention comprehensively consider two dimensions: personal historical behavior and overall community trends when evaluating a user's willingness to break out of the circle.

[0174] Next, obtaining the manager's decision score; including:

[0175] The tags of the third short video recommendation list are used as input to the interest description big language model to generate a text description of community interests; and the tags of the fourth short video recommendation list are used as input to the breaking circle strategy description big language model to generate a text description of the breaking circle strategy.

[0176] It should be noted that the interest description large language model and the breaking circle strategy description large language model here are both trained using the Transformer deep learning architecture. Since the training process itself is an existing technology, it will not be repeated here.

[0177] Based on the text description of the community interests and the text description of the breaking circle strategy, combined with the evaluation results given by the manager, the manager's decision score is defined as S decide , the value is between 0 and 1, and aims to intuitively reflect the manager's comprehensive ability and effectiveness in balancing the interests of all parties, grasping market dynamics and guiding platform development.

[0178] For example, the evaluation given by the manager can be based on multiple dimensions, such as the feasibility of the service for various types of users, the demand for the service in the target market, the application demand of the service from the company itself, and whether there is sufficient financial support for the service. A score of 0-25 is given for each dimension based on the evaluation results, and the total score is normalized to obtain the manager's decision score S decide .

[0179] It is understandable that when formulating a video recommendation strategy, the embodiments of the present invention recognize that it is also crucial to receive evaluation results from managers to ensure that the strategy meets social and platform needs and optimizes the recommendation effect.

[0180] Further, based on the target user's willingness to break the circle and the manager's decision score, the proportion of the fourth short video recommendation list is calculated, and combined with the third short video recommendation list, a final short video recommendation list is obtained; including:

[0181] Calculate the short video recommendation list V Recommend The number of short videos is:

[0182] NUM(V last ))=γ1NUM(V UCF )+γ2NUM(V breakoout ) (8)

[0183]

[0184] Among them, γ1, γ2 represent the corresponding weight coefficients, and γ1+γ2=1, γ1, γ2∈(0,1); NUM(V UCF ),NUM(V breakout ) represent the third short video recommendation list V UCF And the fourth short video recommendation list V breakout The number of short videos;

[0185] Select V according to the ratio UCF and V breakout The short videos with the highest to lowest ratings are combined to obtain the final recommendation list V Recommend .

[0186] In step S7, a pre-trained large language model for title generation based on a Transformer deep learning architecture is used to generate a title of each short video in the short video recommendation list that originates from the fourth short video recommendation list.

[0187] In order to assist the target users to break the circle to the greatest extent, this step generates a large language model based on the title for each short video in the short video recommendation list that originates from the fourth short video recommendation list, tailors an attractive title, captures the user's interest focus and emotional resonance points, so as to stimulate the user's desire to click and interest in watching, thereby improving the attractiveness and conversion rate of the recommended content and attracting users to interact with the recommended videos.

[0188] So far, the embodiment of the present invention has completed the entire process of the short video recommendation method based on the large language model.

[0189] Embodiment 2:

[0190] The embodiment of the present invention provides a short video recommendation system based on a large language model, including:

[0191] An acquisition module, used to acquire a short video set and a user set including a target user;

[0192] The recognition module is used to generate a description text of each short video using a pre-trained BERT-based short video content generation large language model, and extract keywords from the description text using a TF-IDF algorithm to obtain a number of tags for each short video, and obtain a tag set of the short video set after optimization and merging;

[0193] and for counting the historical short video sets of interest of each user within a preset historical time period, combining the tag set to obtain a number of interest tags of each user, and clustering using the K-means algorithm to determine the interest community to which each user belongs;

[0194] The recommendation module is used to compare the similarity between the interest tag of the target user and the tag of any short video, and generate a first short video recommendation list based on the user's personal interest; in the interest community to which the target user belongs, a collaborative filtering method based on neighboring users is used to generate a second short video recommendation list based on the interest community; and a Boosting algorithm is used to fuse the first short video recommendation list and the second short video recommendation list to obtain a third short video recommendation list based on interest;

[0195] and for respectively locating the communities closest to and farthest from the interest community to which the target user belongs, and constructing corresponding candidate short video recommendation lists; integrating the candidate short video recommendation lists of the nearest and farthest communities, and allocating the number of recommended short videos based on the distance, so as to obtain a fourth short video recommendation list that assists in breaking the circle;

[0196] The optimization module is used to calculate the proportion of the fourth short video recommendation list based on the target user's willingness to break the circle and the manager's decision score, and obtain the final short video recommendation list in combination with the third short video recommendation list; wherein the pre-trained interest description large language model based on the Transformer deep learning architecture and the breaking circle strategy description large language model are used, combined with the evaluation results given by the manager, to obtain the manager's decision score.

[0197] Embodiment 3:

[0198] An embodiment of the present invention provides a storage medium storing a computer program for short video recommendation based on a large language model, wherein the computer program enables a computer to execute the short video recommendation method as described in Example 1.

[0199] Embodiment 4:

[0200] An embodiment of the present invention provides an electronic device, including:

[0201] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include a program for executing the short video recommendation method as described in Example 1.

[0202] It is understandable that the short video recommendation system, storage medium and electronic device based on the large language model provided in the embodiments of the present invention correspond to the short video recommendation method based on the large language model provided in the embodiments of the present invention. The explanations, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the short video recommendation method and will not be repeated here.

[0203] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0204] 1. The embodiment of the present invention integrates video content to generate a big language model, an interest description big language model, and a breaking circle strategy description big language model, designs a recommendation system architecture to assist users in breaking the circle, and provides high-quality and diversified short video recommendations for target users while realizing personalized recommendations, which helps them break out of the information cocoon and contact a wider world.

[0205] 2. The embodiment of the present invention further integrates a large language model for title generation, and tailors attractive titles specifically for the short videos that break the circle in the recommendation list to stimulate users' desire to click and interest in watching, thereby improving the attractiveness and conversion rate of the recommended content, attracting users to interact with the recommended videos, and helping users to break the circle.

[0206] 3. The embodiment of the present invention designs a dual-track video recommendation strategy, which not only focuses on the core interests of the target users, but also screens short video content that suits their preferences through algorithms to ensure the accuracy and satisfaction of the recommendations; at the same time, it explores the interest boundaries of the target users and selects those short videos that may stimulate the curiosity of the target users and guide them to cross the existing interest categories. Such a recommendation mechanism is intended to promote the diversified development of the interests of the target users, encourage the target users to try new things, thereby broadening their horizons and enjoying a more colorful video viewing experience.

[0207] 4. When evaluating a user's willingness to break out of the circle, the embodiment of the present invention comprehensively considers two dimensions: personal historical behavior and overall community trends; and when formulating a video recommendation strategy, the embodiment of the present invention recognizes that receiving the evaluation results of the manager is also crucial to ensure that the strategy meets social and platform needs and optimizes the recommendation effect.

[0208] It should be noted that, in this article, relational terms such as first and second, etc. are only used 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. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0209] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A short video recommendation method based on a large language model, characterized in that: include: Obtain a short video set and a user set containing the target user; A pre-trained BERT-based short video content generation large language model is used to generate a description text for each short video, and the TF-IDF algorithm is used to extract keywords from the description text to obtain a number of tags for each short video, and the tag set of the short video set is obtained after optimization and merging; Count the historical short video sets of interest to each user within a preset historical time period, combine the tag set to obtain several interest tags of each user, and use the K-means algorithm for clustering to determine the interest community to which each user belongs; Compare the similarity between the target user's interest tag and the tag of any short video to generate a first short video recommendation list based on the user's personal interest; in the interest community to which the target user belongs, adopt a collaborative filtering method based on neighboring users to generate a second short video recommendation list based on the interest community; The first short video recommendation list and the second short video recommendation list are merged by using a Boosting algorithm to obtain a third short video recommendation list based on interest; Locate the communities closest to and farthest from the interest community to which the target user belongs, and construct a corresponding candidate short video recommendation list; The candidate short video recommendation lists of the nearest and farthest communities are integrated, and the number of recommended short videos is allocated based on the distance to obtain a fourth short video recommendation list that assists in breaking the circle; Based on the target user's willingness to break the circle and the manager's decision score, the proportion of the fourth short video recommendation list is calculated, and combined with the third short video recommendation list to obtain the final short video recommendation list; wherein the pre-trained interest description large language model based on the Transformer deep learning architecture and the breaking circle strategy description large language model are used, combined with the evaluation results given by the manager, to obtain the manager's decision score.

2. The short video recommendation method according to claim 1, characterized in that: Also includes: A pre-trained large language model for title generation based on a Transformer deep learning architecture is used to generate a title for each short video in the short video recommendation list that originates from the fourth short video recommendation list.

3. The short video recommendation method according to claim 1, characterized in that: The method uses the pre-trained BERT-based short video content to generate a large language model to generate a description text for each short video, and uses the TF-IDF algorithm to extract keywords from the description text to obtain a number of tags for each short video, and optimizes and merges them to obtain a tag set for the short video set; including: For any short video, a motion detection algorithm is used to decode the continuous video stream into a series of key frames f o , where o represents the index of the key frame; Calculate each key frame F o The color histogram defines the color space and color channels, and counts the number of pixels of each color value in each color channel to form the color histogram feature vector v color,o ; The gray-level co-occurrence matrix algorithm is used to extract the texture feature vector v by statistically analyzing the joint probability distribution of pixel pairs in the gray-level image at the preset direction and distance. texture,o ; Use the Canny edge detection algorithm to detect the edges in the image, and combine the Hough transform to extract the shape features to obtain the shape feature set S o , and converted into shape feature vector v shape,o ; Extract the audio features of the short video and convert the dubbing in the short video into a text sequence T audio ; Fusion visual feature vector set {v color,o ,v texture,o ,v shape,o } and the converted text sequence T audio , use the pre-trained BERT-based short video content to generate a large language model for encoding, generate a hidden state vector, and generate a descriptive text D through a fully connected layer or classifier; Preprocessing the descriptive text D, extracting keywords from the preprocessed descriptive text D using a TF-IDF algorithm, and using the extracted keywords as candidate tags; Based on the semantic relationship between the candidate tags, a machine learning algorithm is used to optimize the tags to obtain several tags {l1,l2,…,l m }, where m represents the total number of tags of the short video; Merge the labels of all short videos to obtain the label set L of the short video set total .

4. The short video recommendation method according to claim 1, characterized in that: The statistics of each user's historical short video collection of interest within a preset historical time period are combined with the tag set to obtain a number of interest tags for each user, and the K-means algorithm is used for clustering to determine the interest community to which each user belongs; include: The user set is defined as U = {u1,u2,…,u n }, where n represents the total number of users; define user u i The collection of historically interesting short videos within the preset historical time period is where i m Represents user u i The total number of historical short videos of interest; for each short video v j ∈V i , define the tag set of the short video as L total A tag set representing the short video set; For user u i , build the corresponding label count dictionary in, Indicates that label l is in user u i History of Interesting Short Video Collection V i The number of times it appears in the equation; I represents the indicator function. 1 when it is, otherwise 0; Select the tag whose occurrence frequency is greater than the preset threshold as the user u i Merge all users’ interest tags and remove duplicates to obtain the user interest tag set L top ; Clustering using K-means algorithm: Among them, SSE represents the sum of squares of distances from all users to their corresponding cluster centers; K represents the set number of clusters; C c represents the set of users in the cth cluster; μ c is the center of the cth cluster; Taking minimizing SSE as the optimization goal, iteratively update the cluster center and reallocate users to the nearest cluster center to determine user u i Interest group C pi .

5. The short video recommendation method according to claim 4, characterized in that: The step of comparing the similarity between the target user's interest tag and the tag of any short video to generate a first short video recommendation list based on the user's personal interest comprises: For target users Define a feature vector If and only if L top When {k} is the user's interest tag, Equal to 1; for short video v j , define its label as If and only if L top When {k} is the tag of the video, is equal to 1; where |L top | represents the number of tags in the user’s interest tag set, L top {k} represents the kth tag in the user's interest tag set; The following formula is used to compare the similarity between the target user's interest tag and the tag of any short video: Where sim represents the similarity function, · represents the dot product, and ||·|| represents the Euclidean length of the vector; By comparing the similarities, we find a collection of videos that are similar to the target user’s interests and sort them according to the similarity scores to generate a first short video recommendation list V based on the user’s personal interests. CBR ; and / or The method of using a collaborative filtering method based on neighboring users in the interest community to which the target user belongs to generate a second short video recommendation list based on the interest community; comprising: Calculate the similarity between users and determine the target users The neighbor user set of Utilize the short videos v that each neighbor user in the interest community to which the target user belongs has watched or rated p The target users are predicted by weighted average method. Potential interest in unrated short videos; calculated as follows: in, Indicates the target user For short video v p The predicted score of the target user is N(u), N(u) represents the interest community to which the target user belongs, r u′_vp Represents the neighbor user u′’s response to the short video v p The actual rating of The short videos are sorted based on the predicted scores, and several short videos with the highest scores are selected as the second short video recommendation list V based on the interest community. UCF .

6. The short video recommendation method according to claim 1, characterized in that: The process of obtaining the target user's willingness to break the circle includes: For target users Interest groups Sort the interesting tags of the cluster center, define the first several tags with the most occurrences as the mainstream tags of the community; define Function, calculate the target user The proportion of videos not belonging to the mainstream tag of the community watched during the preset historical period is used as a measure of the target users. An indicator of personal breaking circle activity; Combined with the Jaccard coefficient, define function, as a measure of target users Interest groups Indicators of the tendency of social groups to break out of their circles; Among them, the target users Interest groups The number of mainstream tags is used as the base number; interest communities will be counted in two adjacent preset historical time periods. The change of tags of watching short videos is taken as the change number; The function is the ratio of the difference between the base number and the variation to the sum of the base number and the variation; Define the target user The score function of willingness to break the circle: in, is a value between 0 and 1; α, β represent the corresponding weight coefficients, and α+β=1, α, β∈(0,1); and / or The process of obtaining the manager's decision score includes: Using the tags of the third short video recommendation list as input to the interest description large language model to generate a text description of the community interest; and using the tags of the fourth short video recommendation list as input to the breaking circle strategy description large language model to generate a text description of the breaking circle strategy; Based on the text description of the community interests and the text description of the breaking circle strategy, combined with the evaluation results given by the manager, the manager's decision score is defined as S decide , the value is between 0 and 1.

7. The short video recommendation method according to claim 6, characterized in that: Based on the target user's willingness to break the circle and the manager's decision score, the proportion of the fourth short video recommendation list is calculated, and combined with the third short video recommendation list, a final short video recommendation list is obtained; including: Calculate the short video recommendation list V Recommend The number of short videos is: NUM(V last )=γ1NUM(V UCF )+γ2NUM(V breakout ) Among them, γ1, γ2 represent the corresponding weight coefficients, and γ1+γ2=1, γ1, γ2∈(0,1); NUM(V UCF ),NUM(V breakout ) represent the third short video recommendation list V UCF And the fourth short video recommendation list V breakout The number of short videos; Select V according to the ratio UCF and V breakout The short videos with the highest to lowest ratings are combined to obtain the final recommendation list V Recommend .

8. A short video recommendation system based on a large language model, characterized in that: include: An acquisition module, used to acquire a short video set and a user set including a target user; The recognition module is used to generate a description text of each short video using a pre-trained BERT-based short video content generation large language model, and extract keywords from the description text using a TF-IDF algorithm to obtain a number of tags for each short video, and obtain a tag set of the short video set after optimization and merging; and for counting the historical short video sets of interest of each user within a preset historical time period, combining the tag set to obtain a number of interest tags of each user, and clustering using the K-means algorithm to determine the interest community to which each user belongs; A recommendation module, used to compare the similarity between the interest tag of the target user and the tag of any short video, and generate a first short video recommendation list based on the user's personal interests; In the interest community to which the target user belongs, a collaborative filtering method based on neighboring users is used to generate a second short video recommendation list based on the interest community; The first short video recommendation list and the second short video recommendation list are merged by using a Boosting algorithm to obtain a third short video recommendation list based on interest; and for locating the communities closest to and farthest from the interest community to which the target user belongs, and constructing a corresponding candidate short video recommendation list; The candidate short video recommendation lists of the nearest and farthest communities are integrated, and the number of recommended short videos is allocated based on the distance to obtain a fourth short video recommendation list that assists in breaking the circle; The optimization module is used to calculate the proportion of the fourth short video recommendation list based on the target user's willingness to break the circle and the manager's decision score, and obtain the final short video recommendation list in combination with the third short video recommendation list; wherein the pre-trained interest description large language model based on the Transformer deep learning architecture and the breaking circle strategy description large language model are used, combined with the evaluation results given by the manager, to obtain the manager's decision score.

9. A storage medium, characterized in that: The computer program for recommending short videos based on a large language model is stored therein, wherein the computer program enables a computer to execute the short video recommendation method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: include: one or more processors; Memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include programs for executing the short video recommendation method according to any one of claims 1 to 7.

Images