Short video recommendation method and system based on user data

Abstract

The present application belongs to the technical field of short video recommendation, and discloses a short video recommendation method and system based on user data. The method comprises the following steps: grabbing all short video data of a short video platform and clustering; collecting real-time user data, generating corresponding real-time user portraits, and matching a target short video theme clustering cluster of a target short video theme; generating corresponding real-time recommendation strategies according to the real-time user portraits and the target short video theme, and generating a real-time short video recommendation list; generating corresponding real-time recommendation content according to the real-time user portraits, the real-time recommendation strategies and the real-time short video recommendation list, and sending the real-time recommendation content to a user; and selecting target short video data to send to the user according to real-time selection information and the real-time user portrait of the user. The present application solves the problems of insufficient personalization, poor real-time performance, insufficient content diversity and low intelligentization degree in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of short video recommendation technology, specifically relating to a short video recommendation method and system based on user data. Background Technology

[0002] With the rapid development of the internet, short videos have become an indispensable part of people's daily lives. Implementing short video recommendations on short video platforms is of great significance for improving user experience, increasing user retention, promoting content discovery, enhancing platform competitiveness, increasing commercial value, building social networks and communities, incentivizing content creators, and addressing information overload. However, facing massive amounts of short video data, how to recommend content that matches users' interests and needs has become a pressing issue.

[0003] In the field of short video recommendation, although various recommendation systems have been widely used, existing technologies still have many shortcomings that limit their performance and user experience. The main shortcomings of existing technologies are as follows:

[0004] 1) Lack of personalization: Existing recommendation technologies often use general recommendation algorithms, which cannot deeply explore users' personalized needs and lack the ability to capture and analyze users' real-time information, resulting in recommended content that does not match the user's current interests;

[0005] 2) Poor real-time performance: Existing recommendation technologies have a low update frequency and cannot reflect users' new interests and preferences in real time. Insufficient real-time data processing capabilities lead to recommendation lag and affect user satisfaction.

[0006] 3) Insufficient content diversity: Existing recommendation technologies are prone to getting trapped in "information cocoons," that is, only recommending similar content that users have liked in the past, lacking diversity, making it difficult to balance the recommendation of popular content and long-tail content, resulting in some high-quality content being overlooked;

[0007] 4) Low level of intelligence: Existing recommendation technology lacks a dynamic generation mechanism for recommendation strategies, which makes it impossible for recommendation strategies to meet the dynamic changes of short video platforms and user behavior. Summary of the Invention

[0008] To address the shortcomings of existing technologies, such as insufficient personalization, poor real-time performance, insufficient content diversity, and low level of intelligence, the present invention aims to provide a short video recommendation method and system based on user data.

[0009] The technical solution adopted in this invention is as follows:

[0010] A short video recommendation method based on user data includes the following steps:

[0011] Capture all short video data from short video platforms, and cluster the short video data according to several preset short video themes to obtain short video theme clusters.

[0012] Collect real-time user data, use a user profile generation model to generate corresponding real-time user profiles, and match target short video theme clusters based on the real-time user profiles.

[0013] Based on real-time user profiles and target short video themes, a recommendation strategy generation model is used to generate corresponding real-time recommendation strategies, and clusters are generated based on target short video themes to create a real-time short video recommendation list.

[0014] Based on real-time user profiles, real-time recommendation strategies, and real-time short video recommendation lists, a recommendation content generation model is used to generate corresponding real-time recommendation content, which is then sent to users.

[0015] Based on the real-time selection information returned by the user, select the target short video data corresponding to the real-time short video recommendation list and send it to the user, and update the user's real-time user profile based on the real-time selection information.

[0016] Furthermore, all short video data from short video platforms is crawled, and based on several preset short video themes, the short video data is clustered to obtain short video theme clusters, including the following steps:

[0017] Connect to the data server of the short video platform, capture all the short video data stored on the data server, and preprocess the short video data to obtain several preprocessed short video data.

[0018] A keyword generation model is used to generate a real-time keyword sequence for each preprocessed short video data, and the real-time keyword sequence is used as the real-time retrieval tag for the corresponding preprocessed short video data.

[0019] Using real-time search tags as clustering objects, several short video data with real-time search tags are clustered according to several short video themes to obtain short video theme clusters.

[0020] Furthermore, the keyword generation model is constructed based on the FPN-MLP-CRF algorithm;

[0021] The user profile generation model is built based on the RF-Attention-cGAN algorithm;

[0022] The recommendation strategy generation model is built based on the MPO-MOGRPO algorithm;

[0023] The recommended content generation model is built based on the GCN-LSTM-AIGC algorithm.

[0024] Furthermore, a keyword generation model is used to generate a real-time keyword sequence for each preprocessed short video data, and the real-time keyword sequence is used as the real-time retrieval tag for the corresponding preprocessed short video data, including the following steps:

[0025] Each preprocessed short video data is segmented into frames to obtain several corresponding short video image data, and these short video image data are then input into the keyword generation model.

[0026] A keyword generation model is used to extract real-time image features from each short video image data.

[0027] A keyword generation model is used to convert real-time image features into real-time semantic features;

[0028] Using a keyword generation model, a real-time keyword sequence for the preprocessed short video data is generated based on the real-time semantic features of all short video image data.

[0029] Traverse all preprocessed short video data to obtain the real-time keyword sequence for each preprocessed short video data, and use the real-time keyword sequence as the real-time retrieval tag for the corresponding preprocessed short video data.

[0030] Furthermore, using real-time search tags as clustering objects, several short video data sets with real-time search tags are clustered according to several short video themes to obtain short video theme clusters, including the following steps:

[0031] Using real-time retrieval tags as clustering objects, the AP clustering algorithm is used to cluster the real-time retrieval tags of several short video data, resulting in several retrieval tag cluster centers and corresponding retrieval tag clusters.

[0032] Set a corresponding short video theme for each search tag cluster center, and spread the short video theme to the corresponding search tag cluster cluster to obtain search tag clusters with short video themes set.

[0033] Based on the clustering of search tags with short video themes, several short video data sets with real-time search tags are divided into several short video theme clusters.

[0034] Furthermore, real-time user data is collected, and a user profile generation model is used to generate corresponding real-time user profiles. Based on these real-time user profiles, target short video theme clusters are matched, including the following steps:

[0035] Collect real-time user data, preprocess the real-time user data to obtain preprocessed real-time user data, and input the preprocessed real-time user data into the user profile generation model;

[0036] Using a user profiling generation model, several key real-time features of preprocessed real-time user data are extracted.

[0037] Based on the preset attention weight values, the user profile generation model is used to perform weighted fusion of several real-time key features to obtain real-time weighted fusion features.

[0038] Based on the real-time weighted fusion features, a user profile generation model is used to generate corresponding real-time user profiles.

[0039] Obtain the similarity between real-time user profiles and all short video topics, and take the short video topic with the highest similarity as the target short video topic, and extract the target short video topic cluster.

[0040] Furthermore, based on real-time user profiles and target short video themes, a recommendation strategy generation model is used to generate corresponding real-time recommendation strategies, and a real-time short video recommendation list is generated based on target short video theme clustering, including the following steps:

[0041] Based on the target short video theme, the policy network of the recommendation policy generation model is updated to obtain the updated policy network.

[0042] Based on real-time user profiles, a recommendation strategy generation model with an updated policy network is used to generate corresponding real-time recommendation strategies.

[0043] Based on the real-time short video pre-selection decision of the real-time recommendation strategy, several real-time retrieval tags are pre-selected in the target short video topic cluster, and a real-time short video recommendation list is generated based on the several real-time retrieval tags.

[0044] Furthermore, based on real-time user profiles, real-time recommendation strategies, and real-time short video recommendation lists, a recommendation content generation model is used to generate corresponding real-time recommendation content, which is then sent to the user. This includes the following steps:

[0045] The recommendation content generation model is used to extract real-time text features of the real-time short video recommendation list;

[0046] The recommendation content generation model is used to extract real-time graph structure features of real-time user profiles;

[0047] Based on the real-time recommendation content decision, real-time text features, and real-time graph structure features of the real-time recommendation strategy, a recommendation content generation model is used to generate corresponding real-time video recommendation content.

[0048] Based on the real-time content recommendation decision made by the real-time recommendation strategy, real-time recommended content is sent to the user.

[0049] Furthermore, based on the real-time selection information returned by the user, the corresponding target short video data from the real-time short video recommendation list is selected and sent to the user. The user's real-time user profile is then updated based on the real-time selection information, including the following steps:

[0050] Based on the real-time selection information returned by the user, the corresponding target real-time search tags are matched in the real-time short video recommendation list;

[0051] Extract the target short video data corresponding to the target real-time search tag, and send the target short video data to the user based on the real-time short video push decision of the real-time recommendation strategy;

[0052] Based on real-time selection information, corresponding real-time user preference data is generated, and the real-time user profile is updated based on the real-time user preference data, and the updated real-time user profile is obtained and saved.

[0053] A short video recommendation system based on user data is provided to implement a short video recommendation method. The system includes a short video data clustering unit, a user profile generation unit, a recommendation strategy generation unit, a recommendation content generation unit, and a short video data push unit connected in sequence.

[0054] The beneficial effects of this invention are as follows:

[0055] This invention discloses a short video recommendation method and system based on user data. It generates corresponding real-time user profiles based on real-time user data and recommends short videos based on these profiles. This deeply mines and analyzes users' personalized needs, providing more personalized short video content to meet diverse user demands and ensure that recommended content better matches users' current interests and needs. The system updates user profiles based on real-time user selections to ensure that recommended content always aligns with user preferences, further enhancing personalization. Real-time data processing technology is employed to capture and analyze users' latest behaviors, enabling real-time updates of recommended content and significantly improving user experience. By capturing and analyzing all short video data from short video platforms and clustering it with preset short video themes, a balance between popular and long-tail content recommendations is achieved, increasing the diversity of recommended content, avoiding the "information cocoon" effect, and allowing more high-quality content to gain exposure. Through dynamic recommendation strategy generation, the system automatically adjusts the recommendation strategy based on dynamic changes in short video platforms and user behavior, ensuring the accuracy and relevance of recommended content, as well as the level of intelligence in the recommendations.

[0056] Other beneficial effects of the present invention will be further explained in the specific embodiments. Attached Figure Description

[0057] Figure 1 This is a flowchart of the short video recommendation method based on user data in this invention.

[0058] Figure 2 This is a structural block diagram of the short video recommendation system based on user data in this invention. Detailed Implementation

[0059] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0060] Example 1:

[0061] like Figure 1 As shown, this embodiment provides a short video recommendation method based on user data, including the following steps:

[0062] S1: Capture all short video data from short video platforms, and cluster the short video data according to several preset short video themes to obtain short video theme clusters, including the following steps:

[0063] S1-1: Connect to the data server of the short video platform, capture all the short video data stored on the data server, and preprocess the short video data to obtain several preprocessed short video data.

[0064] S1-2: Use a keyword generation model to generate a real-time keyword sequence for each preprocessed short video data, and use the real-time keyword sequence as the real-time retrieval tag for the corresponding preprocessed short video data.

[0065] The keyword generation model is built on the Feature Pyramid Networks (FPN)-Multi-Layer Perceptron (MLP)-Conditional Random Field Algorithm (CRF) algorithm. The keyword generation model includes an image feature extraction module based on the FPN algorithm, a semantic feature transformation module based on the MLP algorithm, and a keyword generation module based on the CRF module, which are connected in sequence.

[0066] The image feature extraction module utilizes the top-down and bottom-up paths of FPN to effectively extract image features at different scales, ensuring accurate detection of various details in the image and achieving the fusion of features at different levels, thus enhancing the expressive power of the features. The semantic feature transformation module learns the complex mapping relationship between input features and output features, performs nonlinear transformations on image features that depend on spatial information, and extracts higher-level semantic features. The keyword generation module considers the global information of the sequence, ensuring that the generated keyword sequence is semantically coherent. Inputting semantic features into the keyword generation module allows for modeling the dependencies between these features and generating keyword sequences. The keyword sequences include several key pieces of information from the short video data, providing support for subsequent topic extraction from the short video data.

[0067] Using a keyword generation model, a real-time keyword sequence is generated for each preprocessed short video data, and the real-time keyword sequence is used as the real-time retrieval tag for the corresponding preprocessed short video data. The process includes the following steps:

[0068] S1-2-1: Perform frame segmentation on each preprocessed short video data to obtain several corresponding short video image data, and input the short video image data into the keyword generation model.

[0069] S1-2-2: Image feature extraction module using keyword generation model to extract real-time image features of each short video image data;

[0070] S1-2-3: The semantic feature conversion module using the keyword generation model converts real-time image features into real-time semantic features;

[0071] S1-2-4: The keyword generation module using the keyword generation model generates a real-time keyword sequence for the preprocessed short video data based on the real-time semantic features of all short video image data.

[0072] S1-2-5: Traverse all preprocessed short video data to obtain the real-time keyword sequence of each preprocessed short video data, and use the real-time keyword sequence as the real-time retrieval tag of the corresponding preprocessed short video data.

[0073] S1-3: Using real-time search tags as clustering objects, cluster several short video data sets with real-time search tags according to several short video themes to obtain short video theme clusters, including the following steps:

[0074] S1-3-1: Using real-time retrieval tags as clustering objects, the Affinity-Propagation (AP) clustering algorithm is used to cluster the real-time retrieval tags of several short video data sets, resulting in several retrieval tag cluster centers and corresponding retrieval tag clusters. This includes the following steps:

[0075] S1-3-1-1: Perform AP clustering algorithm initialization to obtain N initial cluster centers; N is the total number of cluster centers;

[0076] S1-3-1-2: Introduce a reward and punishment mechanism to obtain the reputation value of the real-time search tag for each short video data;

[0077] The formula for reputation score is:

[0078]

[0079] In the formula, C i" The current reputation value for each real-time search tag; i" is the real-time search tag indicator; The initial reputation value for each real-time search tag; α1 is the weight of the reward mechanism; R(i") is the reward value for a single real-time search tag; α2 is the weight of the penalty mechanism; P(i") is the penalty value for a single real-time search tag; N is the total number of real-time search tags;

[0080] The formula for the reward and punishment mechanism is:

[0081]

[0082] In the formula, ε represents the number of times the real-time retrieval tag is selected as a cluster center; e is the natural constant;

[0083]

[0084] In the formula, bad i" B represents the number of times a tag is detected in real time; B is the threshold for the number of times a penalty is triggered.

[0085] S1-3-1-3: Based on the reputation value, sort all real-time search tags in descending order of power, and set the optimal bias parameter for the top N real-time search tags;

[0086] S1-3-1-4: Introduce an iterative decay coefficient to update the attractiveness and attribution information of all real-time search tags, and obtain updated attractiveness and attribution information;

[0087] The formula for attraction information is:

[0088]

[0089] In the formula, rt'+1 (i",k") represents the attraction information of real-time search tag k" to real-time search tag i" at iteration number t'+1; a t； (i",j") represents the attribution information of real-time search tag i" to real-time search tag j" at iteration number t'; r t' (i",j") represents the attraction information of real-time search tag j" to real-time search tag i" at iteration number t'; s(i",k") represents the similarity between real-time search tag k" and real-time search tag i"; i",j" and k" are all real-time search tag indicators;

[0090] The formula for attribution information is:

[0091]

[0092] In the formula, r t'+1 (k",k") represents the attractiveness information of the overall suitability of the tag k" as a cluster center when the iteration number t'+1 is retrieved in real time; ∑ j"≠i",k" max{r t'+1 (j",k"),0} represents the attractiveness information of real-time search tag k" to other real-time search tags besides real-time search tag j" at iteration number t'+1; a t'+1 (i",k") represents the attribution information of real-time retrieval tag i" to real-time retrieval tag k" at iteration number t'+1;

[0093] The iterative update formulas for attraction and attribution information are:

[0094] r' t'+1 (i",k")=λ'*r t' (i",k")+(1-λ')*r t'+1 (i",k")

[0095] a' t'+1 (i",k")=λ'*a t (i",k")+(1-λ')*a t'+1 (i",k")

[0096] In the formula, r' t'+1 (i",k")、a' t'+1 (i",k") represents the attractiveness information and updated attribution information of the real-time search tag k" to the real-time search tag i" at iteration number t'+1; λ' is the iteration decay coefficient; r t' (i",k")、r t'+1 (i",k") represents the attraction information of real-time search tag k" to real-time search tag i" at iteration numbers t' and t'+1; a t' (i",k")、a t'+1(i",k") represents the attribution information of real-time retrieval tag k" to real-time retrieval tag i" at iteration times t' and t'+1;

[0097] S1-3-1-5: Based on the updated attraction information and updated affiliation information, update the N initial cluster centers to obtain N updated cluster centers;

[0098] The formula for determining cluster centers is:

[0099] k"=argmax{a(i",k")+r(i",k")}

[0100] In the formula, i" and k" are both real-time retrieval tag indicators; if i" = k, then real-time retrieval tag i" is the cluster center of real-time retrieval tag k"; if i" ≠ k", then real-time retrieval tag k" is not the cluster center of real-time retrieval tag i".

[0101] S1-3-1-6: Use the updated cluster centers as the search tag cluster centers, and obtain the Euclidean distance between each real-time search tag and all search tag cluster centers;

[0102] S1-3-1-7: Divide all real-time search tags into the search tag cluster centers with the closest Euclidean distance, resulting in several search tag cluster clusters;

[0103] S1-3-2: Set a corresponding short video theme for each search tag cluster center, and spread the short video theme to the corresponding search tag cluster cluster to obtain a search tag cluster with a short video theme set.

[0104] S1-3-3: Based on the clustering of search tags with short video themes, divide the data of several short video data with real-time search tags to obtain several short video theme clusters.

[0105] S2: Collect real-time user data, use the user profile generation model to generate corresponding real-time user profiles, and match the target short video theme clusters based on the real-time user profiles.

[0106] The user profile generation model is built on the Random Forest (RF)-Attention-Conditional Generative Adversarial Network (cGAN) algorithm. The user profile generation model includes a key feature extraction module built on the RF algorithm, an attention weight module built on the Attention mechanism, and a user profile generation module built on the cGAN algorithm, which are connected in sequence. The user profile generation module includes a generator and a discriminator connected in sequence. The generator is connected to the attention weight module.

[0107] The key feature extraction module uses an internal Classification and Regression Tree (CART) to filter key features from the input user data, extracting several key features related to user profile tags, including user information, access preferences, and short video preferences. The attention weighting module uses preset attention weights to perform weighted fusion of several key features, enhancing the influence of the attention features on tag prediction and improving the accuracy of tag prediction. The generator's task is to generate the corresponding user profile based on the weighted fusion features output by the attention weighting module. The discriminator's task is to determine whether the user profile output by the generator is real, i.e., whether it matches the real user profile. Through this adversarial training, the generator continuously optimizes its generation capabilities until it can generate high-quality user profiles.

[0108] Collect real-time user data, use a user profile generation model to generate corresponding real-time user profiles, and match target short video theme clusters based on the real-time user profiles, including the following steps:

[0109] S2-1: Collect real-time user data, preprocess the real-time user data to obtain preprocessed real-time user data, and input the preprocessed real-time user data into the user profile generation model.

[0110] S2-2: Key feature extraction module using user profile generation model to extract several real-time key features of preprocessed real-time user data;

[0111] S2-3: Based on the preset attention weight values, the attention weight module of the user profile generation model is used to perform weighted fusion of several real-time key features to obtain real-time weighted fusion features.

[0112] S2-4: Based on the real-time weighted fusion features, use the user profile generation module of the user profile generation model to generate the corresponding real-time user profile;

[0113] S2-5: Obtain the similarity between the real-time user profile and all short video topics, and take the short video topic with the highest similarity as the target short video topic, and extract the target short video topic cluster of the target short video topic;

[0114] S3: Based on real-time user profiles and target short video themes, use a recommendation strategy generation model to generate corresponding real-time recommendation strategies, and generate a real-time short video recommendation list by clustering based on target short video themes.

[0115] The recommendation policy generation model is built on the Meta-Policy Optimization (MPO)-Multi-Objective Group Relative Policy Optimization (MOGRPO) algorithm. The recommendation policy generation model includes a meta-policy optimization module based on the MPO algorithm and a recommendation policy generation module based on the MOGRPO algorithm. The recommendation policy generation module includes an objective function set, an experience replay pool, a policy network, and an agent. The agent is connected to the objective function set, the experience replay pool, and the policy network, respectively. The meta-policy optimization module is connected to the policy network and the experience replay pool, respectively.

[0116] The meta-policy optimization module optimizes the network parameters of the policy network in the recommendation policy generation module, enabling these parameters to quickly adapt to new and unseen short video topics, thus improving the model's generalization ability. Even with unseen short video topics, the policy network can be updated based on previous learning experience, enhancing the adaptability of the recommendation policy generation model. The objective function set of the recommendation policy generation module can handle multiple conflicting optimization objectives, such as recommendation efficiency, recommendation accuracy, and user experience, generating recommendation policies that balance these objectives. The agent learns historical recommendation policies through an experience replay pool, continuously optimizing its own policy generation capabilities. The agent controls the policy network based on the learned experience to generate more effective recommendation policies. The design of the experience replay pool and the agent allows the model to continuously learn and optimize, improving the quality of policy generation. Due to the adoption of a group exploration approach, the recommendation policy generation module can avoid getting trapped in local optima to some extent. The policy network outputs the probability distribution of actions in a given state. The recommendation policy generation module directly updates the policy network through gradients, eliminating the need for the Critic model in traditional reinforcement learning, making the algorithm structure more concise.

[0117] The method for constructing a recommendation strategy generation model includes the following steps:

[0118] A-1: Using the MPO-MOGRPO algorithm, construct an initial recommendation policy generation model; the initial recommendation policy generation model includes an initial meta-policy optimization module and an initial recommendation policy generation module;

[0119] A-2: The optimization objective of the recommendation strategy is taken as the scenario for meta-strategy optimization. Based on several short video themes under different scenarios, the initial meta-strategy optimization module is trained to obtain the final meta-strategy optimization module.

[0120] A-3: Use the final meta-policy optimization module to initialize the policy network of the initial recommendation policy generation module under different scenarios to obtain the initialized policy network;

[0121] A-4: Based on the different scenarios of the meta-policy optimization module, i.e. the optimization goal of the recommendation policy, set the objective function set and experience replay pool for the initial recommendation policy generation module with the initialized policy network, and set the action space and state space for the agent of the initial recommendation policy generation module.

[0122] A-5: Using the recommendation policy generation problem as a simulation environment, and based on the initialized policy network and the agent with action space and state space, an optimized recommendation policy generation module is obtained;

[0123] A-6: Traverse all objective functions in the objective function set, optimize and train the optimized recommendation strategy generation module based on several historical user profiles, obtain the final recommendation strategy generation module, and generate several historical recommendation strategy generation experiences.

[0124] A-7: Integrate the final meta-strategy optimization module and the final recommendation strategy generation module to obtain the final recommendation strategy generation model, and store several historical recommendation strategy generation experiences in the experience replay pool;

[0125] Based on real-time user profiles and target short video themes, a recommendation strategy generation model is used to generate corresponding real-time recommendation strategies. Then, based on target short video themes, clusters are created to generate a real-time short video recommendation list, including the following steps:

[0126] S3-1: Based on the target short video theme, use the meta-policy optimization module of the recommendation strategy generation model to update the policy network of the recommendation strategy generation module, and obtain the updated policy network.

[0127] S3-2: Based on the real-time user profile, the recommendation strategy generation module uses a recommendation strategy generation model with an updated policy network to generate the corresponding real-time recommendation strategy, including the following steps:

[0128] S3-2-1: Analyze the real-time user profile to obtain several real-time user profile states, and update the state space of the agent of the recommendation strategy generation module according to the several real-time user profile states to obtain the updated state space.

[0129] S3-2-2: Randomly select several historical recommendation strategy generation experiences from the experience replay pool of the recommendation strategy generation module, and generate several possible recommendation decision actions based on these historical recommendation strategy generation experiences.

[0130] S3-2-3: Based on several possible recommended decision actions, update the action space of the agent in the recommendation policy generation module to obtain the updated action space;

[0131] S3-2-4: Select a real-time objective function from the set of objective functions in the recommendation strategy generation module, and based on the real-time objective function, use an agent to control the updated policy network to generate the probability distribution of all possible recommendation decision actions in the updated action space corresponding to each real-time user profile state in the updated state space.

[0132] S3-2-5: Take the most probable recommended decision action in the updated action space as the corresponding real-time user profile state and execute the recommended decision action.

[0133] S3-2-6: Integrate the execution recommendation decision actions of all real-time user profile states in the updated state space to obtain the real-time recommendation strategy;

[0134] S3-3: Based on the real-time short video pre-selection decision of the real-time recommendation strategy, pre-select several real-time retrieval tags in the target short video topic cluster, and generate a real-time short video recommendation list based on the several real-time retrieval tags;

[0135] Real-time short video pre-selection decisions are generated based on short video content that users may accept, which are used to determine the list of short videos to recommend to users;

[0136] S4: Based on real-time user profiles, real-time recommendation strategies, and real-time short video recommendation lists, use the recommendation content generation model to generate corresponding real-time recommendation content and send the real-time recommendation content to users;

[0137] The recommendation content generation model is built on the Graph Convolutional Network (GCN)-Long Short-Term Memory (LSTM)-Artificial Intelligence Generated Content (AIGC) algorithm. The recommendation content generation model includes a graph structure feature extraction module based on the GCN algorithm, a text feature extraction module based on the LSTM algorithm, and a recommendation content generation module group based on the AIGC algorithm. The graph structure feature extraction module and the text feature extraction module are both connected to the recommendation content generation module group. The recommendation content generation module group includes a text generation module based on the Transformer algorithm, an image generation module based on Generative Adversarial Networks (GAN), and a video generation module based on Video Generative Adversarial Networks (VGAN).

[0138] The graph structure feature extraction module extracts corresponding graph structure user profile features from the graph structure user profile through convolution. The graph structure features of the user profile include node features of user tags (user information features, access preference features, short video preference features, etc.) and edge features between user tags. The text feature extraction module is used to extract text features from the short video recommendation list to support subsequent recommendation content generation. The AIGC algorithm of the recommendation content generation module refers to the algorithm that uses artificial intelligence technology to generate content. These algorithms usually include natural language processing, deep learning, generative adversarial networks, variational autoencoders, and other technologies to create text, images, videos, and other forms of data or media content. It refines and simplifies all the short video content included in the short video recommendation list to improve recommendation efficiency and user experience.

[0139] Based on real-time user profiles, real-time recommendation strategies, and real-time short video recommendation lists, a recommendation content generation model is used to generate corresponding real-time recommendation content, which is then sent to users. This process includes the following steps:

[0140] S4-1: Text feature extraction module using the recommendation content generation model to extract real-time text features of the real-time short video recommendation list;

[0141] S4-2: The graph structure feature extraction module of the recommendation content generation model extracts real-time graph structure features of real-time user profiles;

[0142] S4-3: Based on the real-time recommendation content decision of the real-time recommendation strategy, call the text generation module, image generation module and / or video generation module of the recommendation content generation module group of the recommendation content generation model;

[0143] S4-4: Based on real-time text features and real-time graph structure features, use the called text generation module, image generation module and / or video generation module to generate corresponding real-time text recommendation content, real-time image recommendation content and / or real-time video recommendation content;

[0144] Real-time content recommendation decisions are generated based on users' access preferences, short video preferences, and acceptance of recommended content formats. These decisions determine the presentation format of recommended content, including text, images, and / or videos.

[0145] S4-5: Based on the real-time content recommendation decision of the real-time recommendation strategy, send real-time recommended content, including real-time text recommendation content, real-time image recommendation content and / or real-time video recommendation content, to the user;

[0146] Real-time content recommendation decisions are generated based on the user's experience of receiving recommended content, which are used to determine the time period and frequency of sending recommended content;

[0147] S5: Based on the real-time selection information returned by the user, select the target short video data corresponding to the real-time short video recommendation list and send it to the user. Then, based on the real-time selection information, update the user's real-time user profile, including the following steps:

[0148] S5-1: Based on the real-time selection information returned by the user, match the corresponding target real-time search tags in the real-time short video recommendation list;

[0149] S5-2: Extract the target short video data corresponding to the target real-time search tag, and send the target short video data to the user based on the real-time short video push decision of the real-time recommendation strategy;

[0150] S5-3: Generate corresponding real-time user preference data based on real-time selection information, update the user's real-time user profile based on the real-time user preference data, and obtain and save the corresponding updated real-time user profile.

[0151] Example 2:

[0152] like Figure 2 As shown, this embodiment provides a short video recommendation system based on user data to implement a short video recommendation method. The system includes a short video data clustering unit, a user profile generation unit, a recommendation strategy generation unit, a recommendation content generation unit, and a short video data push unit connected in sequence.

[0153] The short video data clustering unit is used to capture all short video data from short video platforms and cluster several short video data according to several preset short video themes to obtain short video theme clusters.

[0154] The user profile generation unit is used to collect real-time user data, use the user profile generation model to generate corresponding real-time user profiles, and match the target short video theme clusters based on the real-time user profiles.

[0155] The recommendation strategy generation unit is used to generate corresponding real-time recommendation strategies based on real-time user profiles and target short video themes using the recommendation strategy generation model, and generate a real-time short video recommendation list based on clustering based on target short video themes.

[0156] The recommendation content generation unit is used to generate corresponding real-time recommendation content based on real-time user profiles, real-time recommendation strategies, and real-time short video recommendation lists, using the recommendation content generation model, and then send the real-time recommendation content to users.

[0157] The short video data push unit is used to select target short video data from the real-time short video recommendation list and send it to the user based on the real-time selection information returned by the user, and update the user's real-time user profile based on the real-time selection information.

[0158] This invention discloses a short video recommendation method and system based on user data. It generates corresponding real-time user profiles based on real-time user data and recommends short videos based on these profiles. This deeply mines and analyzes users' personalized needs, providing more personalized short video content to meet diverse user demands and ensure that recommended content better matches users' current interests and needs. The system updates user profiles based on real-time user selections to ensure that recommended content always aligns with user preferences, further enhancing personalization. Real-time data processing technology is employed to capture and analyze users' latest behaviors, enabling real-time updates of recommended content and significantly improving user experience. By capturing and analyzing all short video data from short video platforms and clustering it with preset short video themes, a balance between popular and long-tail content recommendations is achieved, increasing the diversity of recommended content, avoiding the "information cocoon" effect, and allowing more high-quality content to gain exposure. Through dynamic recommendation strategy generation, the system automatically adjusts the recommendation strategy based on dynamic changes in short video platforms and user behavior, ensuring the accuracy and relevance of recommended content, as well as the level of intelligence in the recommendations.

[0159] This invention is not limited to the optional embodiments described above, and anyone can derive other various forms of products based on the inspiration of this invention. The specific embodiments described above should not be construed as limiting the scope of protection of this invention; the scope of protection of this invention should be determined by the claims, and the specification can be used to interpret the claims.

Claims

1. A short video recommendation method based on user data, characterized in that: The method comprises the following steps: grabbing all short video data of a short video platform, and clustering the short video data according to a plurality of preset short video themes to obtain a short video theme clustering cluster of the short video themes, which comprises the following steps: connecting to a data server of the short video platform, grabbing all short video data stored in the data server, and pre-processing the short video data to obtain a plurality of pre-processed short video data; using a keyword generation model to generate a real-time keyword sequence of each pre-processed short video data, and taking the real-time keyword sequence as a real-time search label of the corresponding pre-processed short video data; the keyword generation model is constructed based on an FPN-MLP-CRF algorithm; taking the real-time search label as a clustering object, clustering a plurality of short video data provided with the real-time search label according to a plurality of short video themes to obtain a short video theme clustering cluster of the short video themes; collecting real-time user data of a user, using a user portrait generation model to generate a corresponding real-time user portrait, and matching a target short video theme clustering cluster of a target short video theme according to the real-time user portrait; the user portrait generation model is constructed based on an RF-Attention-cGAN algorithm; using a recommendation strategy generation model according to the real-time user portrait and the target short video theme to generate a corresponding real-time recommendation strategy, and generating a real-time short video recommendation list according to the target short video theme clustering cluster; the recommendation strategy generation model is constructed based on an MPO-MOGRPO algorithm; The method comprises the following steps: updating the strategy network of the recommendation strategy generation model according to the target short video theme to obtain an updated strategy network; using the recommendation strategy generation model provided with the updated strategy network to generate a corresponding real-time recommendation strategy according to the real-time user portrait; selecting a plurality of real-time search labels in the target short video theme clustering cluster according to a real-time short video pre-selection decision of the real-time recommendation strategy, and generating a real-time short video recommendation list according to the plurality of real-time search labels; using a recommendation content generation model according to the real-time user portrait, the real-time recommendation strategy and the real-time short video recommendation list to generate corresponding real-time recommendation content, and sending the real-time recommendation content to the user; the recommendation content generation model is constructed based on a GCN-LSTM-AIGC algorithm; according to the real-time selection information returned by the user, selecting corresponding target short video data in the real-time short video recommendation list to send to the user, and updating the real-time user portrait of the user according to the real-time selection information. 2.The short video recommendation method based on user data according to claim 1, characterized in that: using a keyword generation model to generate a real-time keyword sequence of each pre-processed short video data, and taking the real-time keyword sequence as a real-time search label of the corresponding pre-processed short video data, which comprises the following steps: frame processing each pre-processed short video data to obtain a plurality of corresponding short video image data, and inputting the plurality of short video image data into the keyword generation model; using the keyword generation model to extract real-time image features of each short video image data; using the keyword generation model to convert the real-time image features into real-time semantic features; The keyword generation model is used to generate a real-time keyword sequence of the preprocessed short video data according to real-time semantic features of all short video image data. All preprocessed short video data is traversed to obtain a real-time keyword sequence of each preprocessed short video data, and the real-time keyword sequence is used as a real-time retrieval label of the corresponding preprocessed short video data. 3.The short video recommendation method based on user data according to claim 2, characterized in that: The real-time retrieval label is used as a clustering object, and a plurality of short video data provided with the real-time retrieval label is clustered according to a plurality of short video themes to obtain a short video theme clustering cluster of the plurality of short video themes, including the following steps: The real-time retrieval label is used as a clustering object, and the AP clustering algorithm is used to cluster the real-time retrieval label of the plurality of short video data to obtain a plurality of retrieval label clustering centers and corresponding retrieval label clustering clusters. A corresponding short video theme is set for each retrieval label clustering center, and the short video theme is diffused to the corresponding retrieval label clustering cluster to obtain a retrieval label clustering cluster provided with the short video theme. The plurality of short video data provided with the real-time retrieval label is divided according to the retrieval label clustering cluster provided with the short video theme to obtain a short video theme clustering cluster of the plurality of short video themes.

4. The short video recommendation method based on user data according to claim 3, characterized in that: Real-time user data of a user is collected, a corresponding real-time user portrait is generated using a user portrait generation model, and a target short video theme clustering cluster of a target short video theme is matched according to the real-time user portrait, including the following steps: Real-time user data of a user is collected, preprocessed to obtain preprocessed real-time user data, and input into the user portrait generation model. The user portrait generation model is used to extract a plurality of real-time key features of the preprocessed real-time user data. The user portrait generation model is used to weight and fuse the plurality of real-time key features according to a preset attention weight value to obtain real-time weighted fusion features. The user portrait generation model is used to generate a corresponding real-time user portrait according to the real-time weighted fusion features. The similarity between the real-time user portrait and all short video themes is obtained, the short video theme with the highest similarity is taken as the target short video theme, and a target short video theme clustering cluster of the target short video theme is extracted.

5. The short video recommendation method based on user data according to claim 4, characterized in that: A corresponding real-time recommendation content is generated using a recommendation content generation model according to the real-time user portrait, a real-time recommendation strategy, and a real-time short video recommendation list, and the real-time recommendation content is sent to the user, including the following steps: The recommendation content generation model is used to extract real-time text features of the real-time short video recommendation list. The recommendation content generation model is used to extract real-time graph structure features of the real-time user portrait. The recommendation content generation model is used to generate corresponding real-time video recommendation content according to real-time recommendation content decisions of the real-time recommendation strategy, real-time text features, and real-time graph structure features. The real-time recommendation content is sent to the user according to real-time content recommendation decisions of the real-time recommendation strategy.

6. The short video recommendation method based on user data according to claim 5, characterized in that: According to the real-time selection information returned by the user, the corresponding target short video data in the real-time short video recommendation list is selected and sent to the user, and the real-time user portrait of the user is updated according to the real-time selection information, including the following steps: According to the real-time selection information returned by the user, a corresponding target real-time search label is matched in a real-time short video recommendation list; Target short video data corresponding to the target real-time search label is extracted, and the target short video data is sent to the user according to a real-time short video pushing decision of a real-time recommendation strategy; According to the real-time selection information, corresponding real-time user preference data is generated, and the real-time user portrait of the user is updated according to the real-time user preference data to obtain and save an updated real-time user portrait. 7.A short video recommendation system based on user data, configured to implement the short video recommendation method according to any one of claims 1-6. The system comprises a short video data clustering unit, a user portrait generation unit, a recommendation strategy generation unit, a recommended content generation unit and a short video data pushing unit connected in sequence.

Images