A video platform cold start recommendation method and system

Abstract

The application provides a video platform cold start recommendation method, and belongs to the technical field of video recommendation. The method comprises the following steps: extracting key information of video content from different types of data of each video in a video platform, and determining video tags of each video according to the key information; obtaining interest tags selected by a user who logs in the video platform for the first time from the video tags in a registration process of the video platform; recommending multiple test videos including interest videos and interest unknown videos to the user based on the interest tags; updating the interest tags selected by the user based on operation behaviors of the user on the multiple test videos, and obtaining user interest tags; obtaining a user interest video sequence based on a KGCN improved algorithm according to video tags of all videos, the user interest tags and a pre-constructed tag knowledge graph, and recommending the user interest videos to the user according to the user interest video sequence.

Description

Technical Field

[0001] This invention belongs to the field of video recommendation technology, and specifically relates to a cold start recommendation method and system for video platforms. Background Technology

[0002] With the widespread adoption of mobile smart devices such as smartphones and tablets, and the increasing fragmentation of people's time, a growing number of video-centric information platforms (referred to as video platforms) have emerged in recent years. Video platforms primarily provide users with personalized video recommendation services; however, during the initial login phase, they lack the necessary user information for personalized recommendations, resulting in a serious cold-start recommendation problem.

[0003] Current video cold start recommendation methods have many shortcomings. First, the video features extracted by existing cold start recommendation methods are relatively simple, and the rich text content in the video is often ignored. For example, a cold start recommendation method for new users only uses the video's statistical data as its feature, ignoring text information such as the video's title. Another example is a cold start recommendation method for newly uploaded videos, which extracts the video's feature vector. For videos with the same content but different presentation styles (such as color tone, filters, etc.), their feature vectors differ greatly, causing the recommendation system to focus more on the video's presentation style and ignore the video content.

[0004] Secondly, most existing cold start recommendation methods ignore differences in user interests, resulting in a lack of personalization in the recommendations. For example, one method calculates scores based on video views, comments, and shares, and then ranks them accordingly, recommending videos with higher scores to new users. However, this method ignores the individual differences of new users, generating the same recommendation list for all new users, relying entirely on statistical data without considering individual user differences. Furthermore, most existing cold start recommendation methods have poor interpretability, making it difficult to build user trust in the platform's recommendation strategy. For instance, one method recommends videos with feature vectors similar to those of previously viewed videos, but cannot explain what features led to the recommendation.

[0005] Finally, existing video cold start methods primarily target recommendations after a platform has accumulated a certain amount of user and video interaction data, following the addition of new users or the uploading of new videos. Therefore, most existing cold start recommendation methods require the platform to accumulate a certain amount of user and video interaction data, using this interaction content as prior knowledge for the recommendation model to complete recommendations. This makes them unsuitable for video platforms in the initial stage where there is a complete lack of user and video interaction data. For example, a cold start recommendation method for new users cannot be applied to the initial stage where user and video interaction data is scarce. Another cold start recommendation method uses the Inception network to extract video feature vectors and combines these feature vectors with the user's historical viewing records for video recommendations. Similarly, this method is also unsuitable for the initial stage of a video platform.

[0006] Therefore, an effective cold start recommendation method for video platforms is needed to address the aforementioned shortcomings or problems. Summary of the Invention

[0007] The purpose of this invention is to provide a cold start recommendation method and system for video platforms to solve or mitigate at least one of the problems in the background art.

[0008] The technical solution of this invention is: a cold start recommendation method for a video platform, comprising:

[0009] Extract key information about the video content from different types of data for each video on the video platform, and determine the video tag for each video based on the key information;

[0010] Obtain the interest tags selected by users who log in to the video platform for the first time during the registration process;

[0011] Based on the aforementioned interest tags, multiple test videos, including interest-based videos and videos with unknown interests, are recommended to the user.

[0012] The user's selected interest tags are updated based on the user's interaction with multiple test videos to obtain the user's interest tags;

[0013] Based on the video tags of all videos, user interest tags, and a pre-built tag knowledge graph, a user interest video sequence is obtained using an improved algorithm based on KGCN, and user interest videos are recommended to users according to the user interest video sequence.

[0014] In at least one embodiment of this application, the different types of data for each video within the video platform include text data, audio data, and image data;

[0015] The process of extracting key information about the video content from different types of data for each video on the video platform, and determining the video tag for each video based on the key information, includes:

[0016] For the text data, natural language processing methods and knowledge graph tools are used to extract keywords, topic words, and node information from the video's text data. These keywords, topic words, and node information are then combined and deduplicated to obtain the video's text data tags. The text data includes the video's title and description. Keywords are at least one word or phrase that appears multiple times in the video's text data, excluding interjections. Topic words are at least one word or phrase that summarizes the content described by the video's text data.

[0017] For audio data, first convert the audio data in the video into text data, and then obtain the audio data tags of the video according to the processing method of text data;

[0018] For image data, computer vision methods are used to identify targets appearing in the video, detect whether the targets match the node information in the tag knowledge graph, and use the matched nodes as image data tags for the video; wherein, the targets include people and objects, and the tag knowledge graph is a graph database composed of nodes representing tags and the relationships between nodes;

[0019] The text data tags, audio data tags, and image data tags are combined and deduplicated to obtain the video tag for the video, and the correlation between the video and its video tag is set to 1.

[0020] In at least one embodiment of this application, the process of obtaining the interest tags selected by a user who is logging into the video platform for the first time from video tags during the video platform registration process includes:

[0021] During the video platform registration process, users select tags from a tag set based on their interests. These tags are then used as the user's chosen interest tags, and for each selected tag, a weight (w) is assigned to the user's interest in that tag. ik =1, where i represents the index of user u, and k represents the index of the selected interest tag in the tag set.

[0022] In at least one embodiment of this application, the process of updating the user's selected interest tags based on the user's operational behavior on multiple test videos to obtain the user's interest tags includes:

[0023] Detect user actions on each test video and quantify the user's level of interest in the corresponding test video;

[0024] Based on the user's interest weights for the selected interest tags and the user's interest in each test video, the updated interest weights for the user for multiple video tags are obtained based on the interest weight update formula.

[0025] Video tags that are greater than the weight threshold in the updated interest weights will be identified as user interest tags.

[0026] In at least one embodiment of this application, the process of detecting user actions on each test video and quantifying the user's level of interest in the corresponding test video includes:

[0027] For a specific test video, detect the user's various actions on that test video and determine the corresponding action weights;

[0028] Based on all user actions and corresponding action weights on the test video, the user's level of interest in the test video is calculated according to their level of interest.

[0029]

[0030] In the formula, p ij Indicates user u i For test video v j The degree of interest; H represents the set of actions, including the user's playback duration or percentage of the test video, whether to like, comment, and favorite; h represents each action in the set of actions H; α h The behavior weight represents the action behavior h;

[0031] By repeating the above process for each test video, the user's level of interest in each test video can be obtained.

[0032] In at least one embodiment of this application, the interest weight update formula includes a first interest weight update formula and a second interest weight update formula;

[0033] The first update formula for the interest weight is: ;

[0034] in, Indicates user u i For label b in the label set k Initial interest weights; Indicates user u i For label b in the label set k Updated interest weights; p ij Indicates user u i For test video v j Degree of interest; t represents time; β t The decay coefficient representing the degree of label relevance. Where 0 < β t<1, e represents the natural constant, η represents the weight of the decay coefficient decreasing over time, Φ(t,t-1) represents the time length since the last interest weight update, and λ represents the bias value;

[0035] The second update formula for the interest weight is: ;

[0036] Where M represents the number of test videos in which the user interacts with multiple test videos; V M This represents the set of test videos that have generated operational behavior; y jk Indicates test video v j and label b in the label set k The degree of correlation, y jk =0 indicates no correlation; test video v j The video tag is not tag b k ;y jk =1 indicates relevance; test video v j The video tag is tag b k .

[0037] In at least one embodiment of this application, the process of obtaining a user interest video sequence based on the KGCN improved algorithm according to the video tags of all videos, user interest tags, and a pre-constructed tag knowledge graph, and recommending user interest videos to the user according to the user interest video sequence includes:

[0038] The user's level of interest in all videos is calculated based on the user's interest tags and the target correlation matrix representing the relationship between each video and tag in the video set and tag set.

[0039] Based on the user's level of interest in all videos, test videos in which the user has already performed actions are removed from the video collection, and the remaining videos are sorted from highest to lowest according to the user's level of interest.

[0040] From the sorted video queue, select a preset number of videos at the top of the sorted list to obtain a sequence of videos that the user is interested in, and recommend videos that the user is interested in based on the sequence of videos that the user is interested in.

[0041] The video set consists of all videos on the video platform. The target correlation matrix is ​​obtained based on the video set, the tag set, and the pre-trained KGCN improved model. The KGCN improved model is trained on the initial model based on KGCN improvement using a preset objective function and loss function, based on the video tags and tag knowledge graph of all videos.

[0042] In at least one embodiment of this application, the process of calculating the user's level of interest in all videos based on user interest tags and a target correlation matrix representing the association between each video and tag in the video set and tag set includes:

[0043] Determine the user interest weight vector based on user interest tags. Among them, the user interest weight vector ∈ User interest weight vector element w in ik Indicates user u i For label b in the label set k The interest weight, where K represents the total number of tags in the tag set;

[0044] Calculate the target relevance matrix Z and the user interest weight vector. The product of these terms yields a vector representing the user's level of interest in all videos. Wherein, the target correlation matrix Z ∈ N represents the total number of videos in the video set.

[0045] In at least one embodiment of this application, the preset objective function is expressed as: ;

[0046] in, The predicted value representing the correlation between label b and video v; Y represents the set of parameters of the model; Y represents the video tag of each video. After setting the correlation degree between the video and its video tag to 1, the correlation degree between all videos and their corresponding video tags is composed of the correlation degree of all videos and their corresponding video tags; G represents the tag knowledge graph; F(.) represents the functional relationship.

[0047] The preset loss function is:

[0048] ;

[0049] Where v represents a video; V represents a collection of videos; y represents the regularization term; λ represents the weight coefficient of the regularization term; jk Indicates label b k and video v j The true value of the correlation; Indicates label b k and video v j The predicted value of the correlation; J(p,q)=-p×log(q) represents the cross-entropy loss function; j: y jk =1 indicates video v j Satisfying condition y jk =1; Tv This represents the number of tags that are irrelevant to video v, and is the number of negative samples.

[0050] The initial model based on KGCN improvement includes an embedding layer, at least one label knowledge graph convolutional layer, and at least one fully connected layer;

[0051] The embedding layer is used to translate a set of input data consisting of video sequence numbers and tag sequence numbers to obtain video embedding vectors and tag embedding vectors respectively.

[0052] The tag knowledge graph convolutional layer is used to perform graph convolution on the tags that are nodes in the tag knowledge graph based on the video embedding vector, the tag embedding vector and the tag knowledge graph, aggregate the information of its P neighboring nodes to obtain the aggregated information of the node's neighboring nodes about the video, and aggregate the node's own embedding vector information and the aggregated information of its neighboring nodes to obtain the node's convolutional embedding vector.

[0053] The fully connected layer is used to output a predicted value of the correlation between the video and the label based on the video embedding vector and the convolutional embedding vector of the label.

[0054] On the other hand, the present invention provides a cold start recommendation system for video platforms, the system comprising:

[0055] The video tag processing module is used to extract key information of video content from different types of data of each video in the video platform, and determine the video tag of each video based on the key information;

[0056] The interest tag processing module is used to obtain the interest tags selected by users who log in to the video platform for the first time from the video tags during the registration process on the video platform;

[0057] The test video generation module is used to recommend multiple test videos to users based on interest tags, including videos with unknown interests.

[0058] The user interest tag processing module is used to update the user's selected interest tags based on the user's operation behavior on multiple test videos, and obtain the user interest tags.

[0059] The video recommendation module is used to obtain user interest video sequences based on the video tags of all videos, user interest tags, and a pre-built tag knowledge graph, using an improved KGCN algorithm, and then recommends user interest videos to the user based on these sequences.

[0060] To address the cold-start video recommendation scenario for video platforms in the initial stage where user and video interaction data is lacking, this invention combines deep learning and knowledge graph technologies to provide a tag-based knowledge graph-based video cold-start recommendation method, aiming to solve the cold-start problem in the initial stage of video platforms in existing technologies. In this embodiment, user and video interaction information is not required as prior knowledge for the recommendation model. Instead, natural language processing and computer vision technologies are combined to comprehensively extract representative tags from the video content information, which are then used as prior knowledge for the recommendation method proposed in this embodiment, making it applicable to the initial stage of video platforms. Each video's video tags fully reflect the content of the video, and each user's determined interest tags reflect the user's interests and take into account user differences. Furthermore, the recall and re-ranking stages of the recommendation rely on tags to generate the recommendation list, thus the recommendation results have good interpretability, helping video platforms to demonstrate the reasons for recommendations to users and build user trust in the platform. As users continue to interact with the videos, the recommendation method proposed in this embodiment can also update user interest tags in a timely manner, making the recommendation effect more in line with users' real-time needs, improving the user's viewing experience, and increasing user retention. Attached Figure Description

[0061] To more clearly illustrate the technical solution provided by the present invention, the accompanying drawings will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention.

[0062] Figure 1 This is a schematic diagram of the video cold start recommendation method of the present invention;

[0063] Figure 2 This is a schematic diagram of a tag knowledge graph in one embodiment of the present invention;

[0064] Figure 3 This is a schematic diagram illustrating the process of extracting video tags from a video in this invention;

[0065] Figure 4 This is a schematic diagram illustrating the process of obtaining user interest tags in this invention;

[0066] Figure 5 This is a schematic diagram illustrating the process of training the initial model based on KGCN improvement in this invention.

[0067] Figure 6 This is a schematic diagram of the video platform cold start recommendation system of this application. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be described in more detail below with reference to the accompanying drawings.

[0069] In view of the cold-start video recommendation scenario in the initial stage of a video platform when there is a lack of user and video interaction data, the present invention combines related technologies such as deep learning and knowledge graphs to provide a cold-start recommendation method and system for a video platform, aiming to solve the cold-start problem in the initial stage of the existing video platform technology.

[0070] As Figure 1 shown, the cold-start recommendation method for a video platform provided by the present invention includes the following steps:

[0071] Step S1, extract the key information of the video content from different types of data of each video in the video platform, and determine the video label of each video according to this key information.

[0072] In order to facilitate the subsequent recommendation of videos that users are interested in, it is first necessary to determine the content information of each video in the video platform before the user selects interest tags. Therefore, before recommending videos, the key information of the video content is extracted in advance from different types of data of each video in the video platform, and the label of the video is determined according to the key information.

[0073] In this application, different types of data of each video include text data, audio data, image data, etc. The process of determining the video label according to the key information of each video content in the video platform is as follows:

[0074] (1) For text data, use natural language processing methods and knowledge graph-related tools to extract keywords, topic words, and node information from the text data of the video, and perform a combined deduplication process on the three to obtain the text data label of the video.

[0075] Among them, the text data includes the title data and summary data of the video. Of course, it can also include the subtitle data of the video, etc.; the keyword is at least one word or phrase that appears multiple times in the text data of the video except for modal particles; modal particles such as "ma", "ba", etc.; the topic word is at least one word or phrase that has a summarizing function for the content described in the text data of the video; the node information refers to at least one word or phrase that matches the node representing the label in the pre-constructed label knowledge graph. The keyword, topic word, and node information may be the same or different.

[0076] Among them, the label knowledge graph is a graph database composed of nodes representing labels and the relationships between nodes, containing the association relationships between labels in the label set. As Figure 2The diagram illustrates the tag knowledge graph relationships between "Player A" and other tags related to the tag "Player A" in one embodiment of the present invention. Circles represent nodes, each node being a representation of a tag, corresponding one-to-one with the tag. Lines connecting nodes represent relationships between them. For example, the relationship between nodes "Player A" and "Team B" is one of "employment," and "Player A" and "Player C" are teammates. The tag knowledge graph can be represented by G, which consists of several triples: head entity-relationship-tail entity, denoted as (b head, r,b tail Examples include ("Player A", "Wife", "Female E"). Please understand the concept of entities in conjunction with existing knowledge graph concepts. In this embodiment of the invention, entities refer to tags. The tag knowledge graph can be pre-constructed based on Wikipedia, books, and other resources, serving as prior knowledge for the recommendation method.

[0077] Tag knowledge graphs can be constructed in advance based on Wikipedia, books, and other resources, serving as prior knowledge for recommendation methods.

[0078] The natural language processing methods and knowledge graph-related tools used can be implemented using existing technologies, such as the Chinese word segmentation tool jieba, the NLTK corpus, and the TF-IDF word frequency scoring algorithm.

[0079] (2) For audio data, first convert the audio data in the video into text data, and then obtain the audio data tags of the video according to the processing method of text data.

[0080] The conversion of audio data into text data can be achieved using existing speech-to-text tools, such as iFlytek speech-to-text cloud service and Tencent Cloud speech recognition. After conversion into text data, keywords, themes and node information are extracted from the text data according to the processing method in (1), and the three are combined and deduplicated to obtain the audio data tags of the video.

[0081] (3) For image data, use computer vision methods to identify the target appearing in the video; detect whether the target matches the node information in the tag knowledge graph, and use the matched node as the image data tag of the video.

[0082] The targets include people and objects. Computer vision methods can be implemented using existing algorithms or models such as object detection, face recognition, and object recognition, including YOLO series networks like YOLOv7 and OpenCV. For example, assuming a face recognition model is used, it can identify whether the target "Player A" is present in the video image data. If so, the face recognition model will output the word "Player A". Then, based on the output of the face recognition model, it can query the label knowledge graph to see if "Player A" exists. If it does, it means that the detected target matches the node information "Player A" in the label knowledge graph, and "Player A" is then used as the image data label for that video.

[0083] Finally, the text data tags, audio data tags, and image data tags are combined and deduplicated to obtain the video tag for the video, and the correlation between the video and its video tag is set to 1.

[0084] Each video's video tag must include at least one word or phrase, and there is no limit to the number of tags that can be repeated.

[0085] For example, in this embodiment of the invention, all videos constitute a video set. Each video in the video platform carries a video tag obtained after the above processing. All video tags constitute a tag set, and each video in the video set is identified by a different sequence number. Similarly, each tag in the tag set is identified by a different sequence number. For ease of understanding, the video set is represented as V = {v1, v2, ..., v...}. j ,...,v N}, where each element v represents a video, and N represents the total number of videos; the tag set is represented as B={b1,b2,...,b k ,...,b K}, where each element b represents a tag, i.e., a video tag, and K represents the total number of tags. The correlation between the video set and the tag set is represented as a matrix. The element y in the j-th row and k-th column of matrix Y jk This refers to video v in the video collection. j With label b in the label set k The degree of relevance, i.e., whether element vj is related to element b. k Related, y jk =0 indicates no correlation, meaning video v j The video tag is not tag b k ;y jk =1 indicates relevance, meaning video v j The video tag is the tag bk.

[0086] In obtaining a video v j Video tag bk Then, the video v j and its video tag b k The correlation is set to 1, which means that the element y in the j-th row and k-th column of matrix Y is... jk Set to 1. It's understandable that if tag b... k It is through the above steps that the video v j The extracted tags, then y jk =1, otherwise y jk =0, therefore, after extracting the video tags for each video, matrix Y contains a large number of zero elements.

[0087] See Figure 3 Understanding the process of extracting video tags from videos according to embodiments of the present invention, as can be seen from the above description, embodiments of the present invention combine technologies such as natural language processing and computer vision to extract representative tags from various aspects of content information such as text data, audio data, and image data of videos. The resulting video tags can fully reflect the content of the video, are more comprehensive, and have higher accuracy.

[0088] S2 retrieves the interest tags selected by users who are logging into the video platform for the first time during the registration process from video tags.

[0089] To provide users with a good personalized recommendation service when they first enter a video platform, they are usually asked to select some tags of interest from a series of labels when registering a video platform account. These selected tags reflect the user's interests, such as music, sports, photography, games, etc. In this way, multiple user-selected interest tags can be obtained.

[0090] Specifically, it obtains the interest tags selected by users during the video platform registration process, including:

[0091] During the video platform registration process, users select tags from a tag set based on their interests. These tags are then used as the user's chosen interest tags, and for each selected tag, a weight (w) is assigned to the user's interest in that tag. ik =1.

[0092] The tag set consists of video tags extracted from all videos on the video platform. For ease of understanding, the user is represented by u, i represents the index of user u, and k represents the index of the selected interest tag in the tag set. Both i and k are natural numbers greater than 0.

[0093] For example, user u i Selected label b k This indicates that user u i For label b kInterested in related videos, set user u i Interest tag b for this selection k The interest weight is w ik =1. For example, if user 2u selects the music tag b1, then w is set to 1. 21 =1; if user u2 selects the sports tag b4, then w is set. 24 =1.

[0094] S3 pushes multiple test videos to users based on interest tags, including videos of interest and videos of unknown interest.

[0095] The video tags are pre-generated based on the content of the corresponding video's audio, text, and image data. The video tags for interest videos are related to the user's selected interest tags. That is, interest videos are identified by searching a set of tags using the user's selected interest tags and determining the videos corresponding to the found tags as interest videos. It is understood that interest videos match the user's interests, but upon the user's initial entry into the platform, to obtain more information about the user's interests, in addition to recommending videos related to content the user explicitly expresses interest in (i.e., interest videos), a certain proportion of videos related to content the user is unsure of their interest in are also recommended—that is, videos with unknown interests. These videos with unknown interests can be randomly selected from the video platform. In some embodiments of this application, the number of test videos pushed to the user each time can be fixed, and the ratio of interest videos to videos with unknown interests can be a fixed ratio. For example, each time 20 test videos are pushed to the user, including 16 interest videos and 4 videos with unknown interests. It is understood that the number of test videos and the ratio of interest videos to videos with unknown interests can also be variable.

[0096] S4 updates the user's selected interest tags based on the user's actions on multiple test videos, thus obtaining the user's interest tags.

[0097] Understandably, user actions in response to test videos can indicate their interest in the content; for example, longer viewing times suggest higher interest. Therefore, user interest tags can be determined based on their interactions with test videos. Combining these tags with the interest tags selected during registration, the final user interest tags can be derived.

[0098] In this application, step S4 specifically includes the following steps:

[0099] S41, Detect user actions on each test video and quantify the user's level of interest in the corresponding test video. This includes the following steps:

[0100] S411, For a given test video, detect the user's various actions on that test video and determine the corresponding action weights. These actions may include the user playing the test video to a certain percentage, liking, commenting, and saving it. Of course, actions are not limited to those listed above.

[0101] For ease of description, each operation is denoted by h, the set of all operations is denoted by H, and the weight of operation h is denoted by α. h This indicates that the value of h may differ for different operational behaviors, and similarly, the corresponding behavior weight α... h They may not be the same.

[0102] For example, when the action h is "play", 0 < h ≤ 1 represents the playback ratio of the test video, that is, the completion rate of the test video watched by the user. If the user watched 30% of the test video, then h = 0.3. When the action h is "like", h = 1 indicates liking, and h = 0 indicates not liking. The actions h for "comment" or "favorite" are similar to those for "like", and will not be elaborated further.

[0103] For all operations h, there are Established. The behavior weight α corresponding to each operation behavior. h You can set the parameters as needed under these conditions; no specific restrictions are imposed here.

[0104] S412, Based on all user actions and corresponding action weights on the test video, calculate the user's level of interest in the test video according to the degree of interest;

[0105] The method for calculating the degree of interest is as follows: In the formula, p ij Indicates user u i For test video v j The degree of interest; H represents the set of operational behaviors; h represents each operational behavior in the set of operational behaviors H; α h The behavior weight represents the operation behavior h.

[0106] It can be seen that the level of interest p ij The size of this value depends on each operation h in the set of operations. The user's interest level (pij) can be quantified by calculating pij. i For test video v j The degree of interest, the degree of interest p ij The larger the value, the stronger the user u i For test video v j The higher the level of interest.

[0107] S413, By repeating the above process for the operation behavior of each test video, the user's level of interest in each test video can be obtained.

[0108] S42, based on the user's interest weights for the selected interest tags and the user's interest in each test video, the updated interest weights for the user for multiple video tags are obtained based on the interest weight update formula.

[0109] Quantify user u i For video v j level of interest p ij In order to establish user u i For video v j Interest weight of video tags.

[0110] The preset interest weight update formulas include a first interest weight update formula and a second interest weight update formula. In practice, one can be used. These are explained below.

[0111] 1) The first update formula for interest weights is: ;

[0112] in, Indicates user u i For label b in the label set k Initial interest weights; Indicates user u i For label b in the label set k Updated interest weights; p ij Indicates user u i For test video v j The degree of interest; t represents time.

[0113] Specifically, This indicates the user u before the interest weights were updated. i For label b k The interest weights in this embodiment of the invention can be updated intermittently. For the first interest weight update, if label b... k User u i The interest tags selected during registration, =1, if label b k Not user u i The interest tags selected during registration, =0. For each subsequent update of the interest weights, This represents the interest weight obtained after the last update.

[0114] β tThis represents the decay coefficient of tag relevance. Considering that user interests gradually change over time, the longer the time since the last update of the interest weight, the greater the decay coefficient β. t It should be larger, ;

[0115] Where, 0 < β t <1; e represents the natural constant; η represents the weight of the decay coefficient decreasing over time; Φ(t,t-1) represents the time length since the last interest weight update; λ represents the bias value.

[0116] Once a user has watched a test video and their level of interest in it is determined, the interest weight of the user's video tag for that test video will be updated once according to the first update formula of interest weight. That is, the first update formula of interest weight adopts an update strategy for each test video, thus enabling timely updates.

[0117] 2) The second update formula for interest weight is:

[0118]

[0119] Where M represents the number of test videos in which the user interacts with multiple test videos; V M This represents the set of test videos that have generated operational behavior; y jk Indicates test video v j and label b in the label set k The degree of correlation, y jk =0 indicates no correlation; test video v j The video tag is not tag b k ;y jk =1 indicates relevance; test video v j The video tag is tag b k .

[0120] In practical applications, besides determining the user's selected interest tags during registration, the user's interest weights for the video tags of the M test videos are often updated all at once after the user has watched M test videos. This reduces the frequency of client-server interaction and alleviates server load. Among these, V... M This represents the set of test videos in which users interacted with multiple test videos. To ensure the timeliness of personalized recommendations, the value of M should not be too large, generally set between 10 and 50, such as 15. It is evident that the second update formula for interest weights employs a centralized update strategy across multiple test videos, thus reducing the load pressure.

[0121] In practical use, you can choose either the first update formula for interest weights or the second update formula for interest weights according to your needs.

[0122] S43 identifies the video tags that are greater than the weight threshold among the multiple updated interest weights as user interest tags.

[0123] Understandably, user u i For label b in the label set k Updated interest weight This is a value between [0,1], and its magnitude directly reflects the user's level of interest. For example, let's say user "Zhang San" is interested in the tag "food". A value of 0.8 indicates that "Zhang San" has a level of interest in "food" of 0.8. Since user interest tags indicate a high level of interest, it's possible to... Set a weight threshold, and then... The corresponding video tags are determined as user interest tags. The weight threshold can be determined based on experience, such as 0.3 or 0.25.

[0124] For details on the process of determining user interest tags in embodiments of the present invention, please refer to [link / reference]. Figure 4 .

[0125] S5: Based on the video tags of all videos, user interest tags, and a pre-built tag knowledge graph, the system obtains user interest video sequences using an improved KGCN algorithm, and recommends user interest videos to users based on these sequences.

[0126] After identifying user interest tags using test videos, personalized recommendation services can be provided by combining this with a tag knowledge graph. As mentioned earlier, the tag knowledge graph contains the relationships between tags in a tag set. The tag knowledge graph helps to uncover user interests; for example, if user u is known... i If user u is interested in "player A" and "team B", then we can infer that user u i Players may also be interested in "Player C" and "Player D," who are also members of Team B. Therefore, the tag knowledge graph is crucial for cold start recommendations.

[0127] Existing technologies include several knowledge graph-based recommendation algorithms, such as Knowledge Graph Attention Network (KGAT) and Knowledge Graph Convolutional Networks (KGCN). However, these algorithms require a large amount of user-video interaction information as prior knowledge to train the recommendation model, making them unsuitable for video platforms in the initial stages where user interaction with videos is scarce. Therefore, this invention proposes an improved cold-start recommendation algorithm based on KGCN, enabling it to address the cold-start problem by combining a tag-based knowledge graph to provide personalized recommendation services to users.

[0128] Specifically, step S5 of the present invention includes the following steps:

[0129] S51. Calculate the user's level of interest in all videos based on user interest tags and the target correlation matrix representing the relationship between each video and tag in the video set and tag set.

[0130] The video set consists of all videos on the video platform, and is represented as V{v1, v2, ..., v...} j ,..., v N}

[0131] The target correlation matrix is ​​obtained based on the video set, the tag set, and the pre-trained KGCN improved model. The KGCN improved model is obtained by training the initial model based on KGCN improvement using the video tags and tag knowledge graph of all videos, and using the preset objective function and loss function.

[0132] To facilitate understanding of the scheme, the following will explain the process of obtaining the KGCN improved model and the process of obtaining the target correlation matrix.

[0133] (1) Improved KGCN model

[0134] For the existing tag set B={b 1, b2,...,b k ,...,b K The main objective of the KGCN improvement method in this embodiment of the invention is to predict the correlation between each tag and each video in the video set and tag set. This includes all user interest tags and a tag knowledge graph.

[0135] The predefined objective function is expressed as: ;

[0136] in, The predicted value representing the correlation between label b and video v; Y represents the set of parameters of the model; Y represents the video tag of each video. After setting the correlation degree between the video and its video tag to 1, the correlation degree between all videos and their corresponding video tags is composed of the correlation degree of all videos and their corresponding video tags; G represents the tag knowledge graph; F(.) represents the functional relationship.

[0137] This invention first constructs an initial model based on an improved KGCN, which includes an embedding layer, at least one tag knowledge graph convolutional layer, and at least one fully connected layer. Then, based on the video tags and tag knowledge graphs of all videos, the initial model based on the improved KGCN is trained using a preset objective function and loss function, resulting in a trained improved KGCN model. The following describes the process in conjunction with... Figure 5 The process of training the initial model based on KGCN improvement is explained. Figure 5 An example is given consisting of two label knowledge graph convolutional layers and two fully connected layers. The embedding layer can typically be represented as an embedding layer. Specifically:

[0138] ① The embedding layer is used to translate a set of input data consisting of video sequence number and tag sequence number to obtain video embedding vector and tag embedding vector respectively;

[0139] Let's take any set of input data input into the initial model improved from KGCN as an example. This set of input data is represented as (video sequence number, tag sequence number). The translation of the video sequence number and the tag sequence number is performed separately to avoid confusion.

[0140] After passing through the embedding layer, the video sequence number is translated into a video embedding vector; after passing through the tag sequence number, it is translated into a tag embedding vector. For a detailed explanation of the specific mechanism of the embedding layer, please refer to the relevant technical documentation; it will not be elaborated upon here.

[0141] ② The tag knowledge graph convolutional layer is used to perform graph convolution on the tags that are nodes in the tag knowledge graph based on the video embedding vector, the tag embedding vector and the tag knowledge graph. It aggregates the information of its P neighboring nodes to obtain the aggregated information of the neighboring nodes about the video. It also aggregates the embedding vector information of the node itself and the aggregated information of its neighboring nodes to obtain the convolutional embedding vector of the node.

[0142] For example Figure 5 In order to fully utilize the information in the tag knowledge graph, it is necessary to perform graph convolution on the nodes to aggregate the information of their neighboring nodes. Considering that different relationships between tags have varying importance to the video (e.g., some videos focus more on "Player A's" family relationships, while others focus more on "Player A's" experiences in team B), this invention uses... Let r represent the importance of relation r to video v, where both relation r and video v can be represented by embedding vectors. Represents the video embedding vector; The embedding vector representing relation r; d represents the dimension of the embedding vector; g represents the function that operates on the two vectors, such as the inner product operation. During aggregation, the aggregation information of node b's neighboring nodes with respect to video v is represented as: ;

[0143] in,

[0144] N(b) represents the aggregated information of the neighboring nodes of node b about video v; N(b) represents the set of neighboring nodes of node b. Represents the embedding vector of any neighboring node; This indicates the importance of the relationship r between nodes b and e to video v; The relationship r between nodes b and h represents the importance of video v; e∈N(b) means that the neighboring node e visited belongs to the set of neighboring nodes; h∈N(b) means that the neighboring node h visited belongs to the set of neighboring nodes.

[0145] In a tag knowledge graph, some nodes may have many neighboring nodes, while others may have very few. To facilitate unified parallel processing, this invention sets a sampling constant P, ensuring that each node logically has P neighbors. During sampling, nodes with more than P neighboring nodes will have some neighbors ignored, while the neighbors of nodes with fewer than P neighboring nodes will be sampled repeatedly. Here, P is a natural number greater than 0, such as 4.

[0146] Then, by aggregating the embedding vector information of the node itself and the aggregation information of its neighboring nodes, the embedding vector of node b after convolution in the current label knowledge graph convolutional layer is obtained, i.e.:

[0147]

[0148] in, This represents the embedding vector of node b received by the current tag knowledge graph convolutional layer; This represents the aggregated information of node b's neighboring nodes regarding video v; for example, for the first layer of the label knowledge graph convolutional layer, it is the label embedding vector output by the embedding layer; for the second layer of the label knowledge graph convolutional layer, This is the convolutional embedding vector of the nodes output by the first-layer label knowledge graph convolutional layer; It is a bias term; , ∈ W represents the weight matrix of the tag knowledge graph convolutional layer, W∈ σ(.) is an activation function similar to the ReLU function. Aggregation methods include additive aggregation, concatenation aggregation, and adjacent aggregation.

[0149] It is understandable that, through the above processing, we can obtain the post-convolution embedding vector of the last tag knowledge graph convolutional layer, that is, the post-convolution embedding vector of the tag.

[0150] ③ Fully connected layer, used to output the predicted value of the correlation between video and label based on the convolutional embedding vector of video and label.

[0151] For ease of understanding, combined with Figure 5 The video embedding vector and the label correspond to the video v respectively. j and label b k For example, after two convolutional layers of the label knowledge graph, each node in the label knowledge graph aggregates information from its neighboring nodes. The fully connected layer receives the video v as input. j Embedded vector and label b k The embedding vector is generated after convolution through two convolutional layers of the labeled knowledge graph. .

[0152] The function of a fully connected layer can be abstractly represented as:

[0153] Where f(.) represents prediction; Indicates label b k and video v j The predicted value of the correlation.

[0154] In this invention, for the case of only one fully connected layer and the aggregation method is splicing aggregation, the correlation prediction value is... It can be represented as:

[0155] in, It is a bias term; D∈ σ represents a matrix of model parameters; concat represents concatenation and aggregation; σ(.) is an activation function similar to ReLU.

[0156] Understandably, during training, multiple positive and negative samples can be identified using the video set and the video labels extracted from each video. A positive sample represents a video v. j and label b k The correlation is defined as the degree of relevance, meaning the tag is a video tag for the video, and the true value y represents the degree of relevance between the video and the tag. jk=1; Negative samples indicate that the correlation between the video and the tag is unrelated, meaning the tag is not the video tag for that video. The true value of the correlation between the video and the tag is y. jk =0. Inputting positive and negative samples into the initial model improved based on KGCN yields predicted values ​​for the correlation between videos and labels in each sample. Based on the difference between the actual and predicted values ​​of the correlation between videos and tags, backpropagation is performed using a pre-defined loss function, with the aim of improving the predicted values. Get as close as possible to the true value y jk After repeated iterations, the parameters of the model training can be obtained, that is, the improved KGCN model after training is obtained.

[0157] The preset loss function is expressed as:

[0158]

[0159] Where v represents a video; V represents a collection of videos; y represents the regularization term; λ represents the weight coefficient of the regularization term; jk Indicates label b k and video v j The true value of the correlation; Indicates label b k and video v j The predicted value of the correlation; J(p,q)=-p×log(q) represents the cross-entropy loss function; j: y jk =1 indicates video v j Satisfying condition y jk =1; T v This represents the number of tags that are irrelevant to video v, and is the number of negative samples.

[0160] Negative samples are obtained through negative sampling, which requires that the number of negative samples be roughly equal to the number of positive samples, i.e., T. v =|{j:y jk =1}|, and when sampling tags unrelated to video v, the more popular the tag, the higher the probability of it becoming a negative sample. This is because, compared to other tags, popular tags are associated with more videos. If a video has no direct association with a popular tag, then that video and that popular tag are more likely to be unrelated. Here, the direct association between a video and a tag is defined as: if tag b k From video v j The extracted value is y. jk =1, then the video v is considered to be 1. j and label b k There is a direct correlation.

[0161] For details on the training process of the improved KGCN model, please refer to the existing neural network training process; it will not be described here again.

[0162] (2) Target correlation matrix

[0163] Once the trained KGCN improved model is obtained, inputting any set of input data (video index, tag index) will yield predicted values ​​for the correlation between videos and tags in that set of input data. Since the correlation between videos and tags is relatively stable, each set of input data obtained from the video set and tag set can be sequentially input into the KGCN improved model to calculate its correlation, resulting in the target correlation matrix Z ∈ [missing information - likely a specific matrix or formula]. The element z jk (0≤z jk ≤1) indicates video v j and label b k The relevance is K, where K represents the total number of tags in the tag set and N represents the total number of videos in the video set.

[0164] Therefore, for step S51, the user's interest in all videos is calculated based on the user's interest tags and the target correlation matrix representing the relationship between each video and tag in the video set and tag set. This can specifically include the following steps:

[0165] S511, Determine the user interest weight vector based on user interest tags. .

[0166] Among them, user u i Interests are represented as user interest weight vectors ∈ User interest weight vector element w in ik Indicates user u i For label b in the label set k Interest weight; w ik It is determined during the process of identifying user interest tags.

[0167] S512, Calculate the target relevance matrix Z and the user interest weight vector. The product of these terms yields a vector representing the user's level of interest in all videos. .

[0168] In order to provide users u i Providing personalized recommendation services requires calculating a vector of the user's interest in all videos. The vector of user interest in all videos. element p in ij Indicates user u iFor video v j The level of interest.

[0169] S52, based on the user's level of interest in all videos, remove test videos from the video collection that the user has already interacted with, and sort the remaining videos in descending order of the user's level of interest.

[0170] It is understandable that vectors It can be represented as {p i1 ,p i2 ,…,p ij, …,p iN There are a total of N elements, corresponding to N videos. Since user u... i If multiple test videos have already been pushed to the user, then the user can... i Test videos that have already generated user interaction are removed, and the remaining videos are sorted in descending order of interest level p to obtain a video queue.

[0171] S53: Select a preset number of videos from the sorted video queue to obtain a sequence of videos that the user is interested in, and recommend videos that the user is interested in based on this sequence.

[0172] The preset quantity can be selected as needed, such as 100.

[0173] Once you have a sequence of videos that a user is interested in, you can push them to the user. For example, you can push them in the form of a list, which may contain links to the videos and corresponding descriptions, or you can push the corresponding videos to the user sequentially according to the sequence.

[0174] Specific details regarding KGCN improvement methods will not be repeated here.

[0175] In this embodiment of the invention, since the target relevance matrix Z is only affected by videos and tags, there is no need to retrain the model when user interests change. Therefore, the target relevance matrix Z can be directly stored in the memory of the video platform's server, allowing all users to share the target relevance matrix Z. Each recommendation only requires the target relevance matrix Z and the current user's interest weight vector. Multiplying these directly calculates the vector of the current user's interest in all videos. This significantly shortens the response time of the recommendation system. Furthermore, the target relevance matrix Z reflects the relationship between videos and tags, and it is completely independent of user and video interaction; the model can obtain the target relevance matrix even without users. User interest weight vector It reflects the user's interest weight in tags, can be initialized during user registration, does not depend on user interaction with videos, and is continuously updated and improved as users interact with videos. Therefore, the video cold start recommendation method based on tag knowledge graph proposed in this invention can be applied to the initial stage of video platforms.

[0176] In summary, in practical scenarios, existing videos in the video platform's database can be linked to tags based on their text, audio, and image information and stored in the database. Whenever a user logs into the video platform via a smartphone or other terminal device, the platform pushes test videos to the user based on the interest tags selected during registration, collects the user's viewing behavior, updates the user's interests based on these behaviors, and associates them with tags to determine the user's interest tags. Finally, combining the tag knowledge graph with the obtained video tags and user interest tags, the KGCN improved method is used to rank the videos. Based on the ranking results, a series of videos that the user is most interested in are recommended to the user as a sequence of videos of interest. Since users will generate new behaviors in response to the recommended videos, the user's interest tags can be updated again, entering the next round of recommendations. This continuous updating based on changes in user interests makes the recommendation effect more in line with the user's real-time needs.

[0177] To address the cold start video recommendation scenario for video platforms in the initial stage where user and video interaction data are lacking, this invention combines deep learning and knowledge graph technologies to provide a video cold start recommendation method based on a tag knowledge graph, aiming to solve the cold start problem in the initial stage of video platforms in existing technologies. This invention pre-constructs a tag knowledge graph and extracts video tags from the videos on the video platform that best reflect the key content of the videos, thus determining a tag set. In the actual recommendation process, for users logging into the video platform for the first time, the invention first obtains the interest tags selected by the user during the video platform registration process from the tag set. Next, based on the user's selected interest tags, multiple test videos, including videos of interest and videos with unknown interest, are pushed to the user. Then, the user's selected interest tags are updated by detecting the user's actions on the multiple test videos, thus obtaining the user's interest tags. Finally, based on the video tags of all videos, the user's interest tags, and the tag knowledge graph, an improved KGCN method is used to obtain the sequence of videos of interest to the user and recommend them to the user.

[0178] This invention does not require user-video interaction information as prior knowledge for the recommendation model. Instead, it combines natural language processing and computer vision technologies to comprehensively extract representative tags from the video content information, using these tags as prior knowledge for the recommendation method proposed in this invention. This approach is suitable for the initial stages of video platforms. Each video's video tags fully reflect its content, and each user's determined interest tags reflect their interests and take into account user differences. Furthermore, both the recall and re-ranking stages of the recommendation rely on tags to generate the recommendation list, thus providing good interpretability for the recommendation results. This helps video platforms demonstrate the reasons for recommendations to users, thereby building user trust. As users continuously interact with videos, the recommendation method proposed in this invention can also update user interest tags in a timely manner, making the recommendation effect more in line with users' real-time needs, improving the user's viewing experience, and increasing user retention.

[0179] Finally, as Figure 6 As shown, the present invention also provides a video platform cold start recommendation system 100, which includes:

[0180] The video tag processing module 101 is used to extract key information of video content from different types of data of each video in the video platform, and determine the video tag of each video based on the key information.

[0181] The interest tag processing module 102 is used to obtain the interest tags selected by users who log in to the video platform for the first time from the video tags during the video platform registration process;

[0182] The test video generation module 103 is used to recommend multiple test videos to users based on interest tags, including interest videos and videos with unknown interests.

[0183] User interest tag processing module 104 is used to update the user's selected interest tags based on the user's operational behavior on multiple test videos, thereby obtaining user interest tags; and

[0184] The video recommendation module 105 is used to obtain the user's interest video sequence based on the video tags of all videos, user interest tags, and a pre-built tag knowledge graph, and recommend the user's interest videos to the user based on the user's interest video sequence.

[0185] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A cold start recommendation method for a video platform, characterized in that, include: Extract key information about the video content from different types of data for each video on the video platform, and determine the video tag for each video based on the key information; Obtain the interest tags selected by users who log in to the video platform for the first time during the registration process; Based on the aforementioned interest tags, multiple test videos, including interest-based videos and videos with unknown interests, are recommended to the user. The user's selected interest tags are updated based on the user's interaction with multiple test videos to obtain the user's interest tags; Based on the video tags of all videos, user interest tags, and a pre-built tag knowledge graph, a user interest video sequence is obtained using an improved algorithm based on KGCN, and user interest videos are recommended to users according to the user interest video sequence.

2. The video platform cold start recommendation method as described in claim 1, characterized in that, The different types of data for each video within the video platform include text data, audio data, and image data; The process of extracting key information about the video content from different types of data for each video on the video platform, and determining the video tag for each video based on the key information, includes: For the text data, natural language processing methods and knowledge graph tools are used to extract keywords, topic words, and node information from the video's text data. These keywords, topic words, and node information are then combined and deduplicated to obtain the video's text data tags. The text data includes the video's title and description. Keywords are at least one word or phrase that appears multiple times in the video's text data, excluding interjections. Topic words are at least one word or phrase that summarizes the content described by the video's text data. For audio data, first convert the audio data in the video into text data, and then obtain the audio data tags of the video according to the processing method of text data; For image data, computer vision methods are used to identify targets appearing in the video, detect whether the targets match the node information in the tag knowledge graph, and use the matched nodes as image data tags for the video; wherein, the targets include people and objects, and the tag knowledge graph is a graph database composed of nodes representing tags and the relationships between nodes; The text data tags, audio data tags, and image data tags are combined and deduplicated to obtain the video tag for the video, and the correlation between the video and its video tag is set to 1.

3. The video platform cold start recommendation method as described in claim 1, characterized in that, The process of obtaining the interest tags selected by users who are logging into the video platform for the first time from video tags during the video platform registration process includes: During the video platform registration process, users select tags from a tag set based on their interests. These tags are then used as the user's chosen interest tags, and for each selected tag, a weight (w) is assigned to the user's interest in that tag. ik =1, where i represents the index of user u, and k represents the index of the selected interest tag in the tag set.

4. The video platform cold start recommendation method as described in claim 3, characterized in that, The process of updating the user's selected interest tags based on the user's interaction with multiple test videos to obtain the user's interest tags includes: Detect user actions on each test video and quantify the user's level of interest in the corresponding test video; Based on the user's interest weights for the selected interest tags and the user's interest in each test video, the updated interest weights for the user for multiple video tags are obtained based on the interest weight update formula. Video tags that are greater than the weight threshold in the updated interest weights will be identified as user interest tags.

5. The video platform cold start recommendation method as described in claim 4, characterized in that, The process of detecting user actions on each test video and quantifying the user's level of interest in the corresponding test video includes: For a specific test video, detect the user's various actions on that test video and determine the corresponding action weights; Based on all user actions and corresponding action weights on the test video, the user's level of interest in the test video is calculated according to their level of interest. ； In the formula, p ij Indicates user u i For test video v j The degree of interest; H represents the set of actions, including the user's playback duration or percentage of the test video, whether to like, comment, and favorite; h represents each action in the set of actions H; α h The behavior weight represents the operation behavior h; By repeating the above process for each test video, the user's level of interest in each test video can be obtained.

6. The video platform cold start recommendation method as described in claim 5, characterized in that, The interest weight update formula includes a first interest weight update formula and a second interest weight update formula; The first update formula for the interest weight is: ; in, Indicates user u i For label b in the label set k Initial interest weights; Indicates user u i For label b in the label set k Updated interest weights; p ij Indicates user u i For test video v j Degree of interest; t represents time; β t The decay coefficient representing the degree of label relevance. Where 0 < β t <1, e represents the natural constant, η represents the weight of the decay coefficient decreasing over time, Φ(t,t-1) represents the time length since the last interest weight update, and λ represents the bias value; The second update formula for the interest weight is: ; Where M represents the number of test videos in which the user interacts with multiple test videos; V M This represents the set of test videos that have generated operational behavior; y jk Indicates test video v j and label b in the label set k The degree of correlation, y jk =0 indicates no correlation; test video v j The video tag is not tag b k ;y jk =1 indicates relevance; test video v j The video tag is tag b k .

7. The video platform cold start recommendation method as described in claim 6, characterized in that, The process of obtaining user interest video sequences based on the video tags of all videos, user interest tags, and a pre-constructed tag knowledge graph, using an improved KGCN algorithm, and recommending user interest videos to users based on these sequences includes: The user's level of interest in all videos is calculated based on the user's interest tags and the target correlation matrix representing the relationship between each video and tag in the video set and tag set. Based on the user's level of interest in all videos, test videos in which the user has already performed actions are removed from the video collection, and the remaining videos are sorted from highest to lowest according to the user's level of interest. From the sorted video queue, select a preset number of videos at the top of the sorted list to obtain a sequence of videos that the user is interested in, and recommend videos that the user is interested in based on the sequence of videos that the user is interested in. The video set consists of all videos on the video platform. The target correlation matrix is ​​obtained based on the video set, the tag set, and the pre-trained KGCN improved model. The KGCN improved model is trained on the initial model based on KGCN improvement using a preset objective function and loss function, based on the video tags and tag knowledge graph of all videos.

8. The video platform cold start recommendation method as described in claim 7, characterized in that, The process of calculating the user's interest in all videos based on user interest tags and a target correlation matrix representing the relationship between each video and tag in the video set and tag set includes: Determine the user interest weight vector based on user interest tags. Among them, the user interest weight vector ∈ User interest weight vector element w in ik Indicates user u i For label b in the label set k The interest weight, where K represents the total number of tags in the tag set; Calculate the target relevance matrix Z and the user interest weight vector. The product of these terms yields a vector representing the user's level of interest in all videos. Wherein, the target correlation matrix Z ∈ N represents the total number of videos in the video set.

9. The video platform cold start recommendation method as described in claim 8, characterized in that, The preset objective function is expressed as follows: ; in, The predicted value representing the correlation between label b and video v; Y represents the set of parameters of the model; Y represents the video tag of each video. After setting the correlation degree between the video and its video tag to 1, the correlation degree between all videos and their corresponding video tags is composed of the correlation degree of all videos and their corresponding video tags; G represents the tag knowledge graph; F(.) represents the functional relationship. The preset loss function is: ； Where v represents a video; V represents a collection of videos; y represents the regularization term; λ represents the weight coefficient of the regularization term; jk Indicates label b k and video v j The true value of the correlation; Indicates label b k and video v j The predicted value of the correlation; J(p,q)=-p×log(q) represents the cross-entropy loss function; j: y jk =1 indicates video v j Satisfying condition y jk =1; T v This represents the number of tags that are irrelevant to video v, and is the number of negative samples. The initial model based on KGCN improvement includes an embedding layer, at least one label knowledge graph convolutional layer, and at least one fully connected layer; The embedding layer is used to translate a set of input data consisting of video sequence numbers and tag sequence numbers to obtain video embedding vectors and tag embedding vectors respectively. The tag knowledge graph convolutional layer is used to perform graph convolution on the tags that are nodes in the tag knowledge graph based on the video embedding vector, the tag embedding vector and the tag knowledge graph, aggregate the information of its P neighboring nodes to obtain the aggregated information of the node's neighboring nodes about the video, and aggregate the node's own embedding vector information and the aggregated information of its neighboring nodes to obtain the node's convolutional embedding vector. The fully connected layer is used to output a predicted value of the correlation between the video and the label based on the video embedding vector and the convolutional embedding vector of the label.

10. A cold start recommendation system for a video platform, characterized in that, include: The video tag processing module is used to extract key information of video content from different types of data of each video in the video platform, and determine the video tag of each video based on the key information; The interest tag processing module is used to obtain the interest tags selected by users who log in to the video platform for the first time from the video tags during the registration process on the video platform; The test video generation module is used to recommend multiple test videos to users based on interest tags, including videos with unknown interests. The user interest tag processing module is used to update the user's selected interest tags based on the user's operation behavior on multiple test videos, and obtain the user interest tags. The video recommendation module is used to obtain user interest video sequences based on the video tags of all videos, user interest tags, and a pre-built tag knowledge graph, using an improved KGCN algorithm, and then recommends user interest videos to the user based on these sequences.

Images