Video recommendation method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a video recommendation method and device, electronic equipment and a storage medium, and relates to the technical field of data processing. The specific implementation scheme is as follows: in response to a search term of a target user, performing video retrieval based on the search term to obtain a plurality of videos; selecting a video of which the online time length is not greater than a preset time length from the plurality of videos as a cold start video; in response to at least one cold start video, calculating an interest expectation value of the target user for each cold start video based on the user characteristics of the target user and the description characteristics of each cold start video; based on the interest expectation value of each cold start video, screening the at least one cold start video to obtain at least one first type candidate video; and recommending at least one target recommended video to the target user based on the at least one first-class candidate video. Therefore, according to the scheme of the invention, the exposure rate of the cold start video can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of data processing, and particularly to a video recommendation method, apparatus, electronic device, and storage medium. Background Art

[0002] In related technologies, after a user inputs a search term for a video, various videos matching the search term can be retrieved, and then, based on a video recommendation model, each video is analyzed to determine a recommended video, and the determined recommended video is recommended to the user. Among them, when the video recommendation model determines the video to be recommended, it is affected by interaction data of specified behaviors such as the click-through rate, view count, and like count of the video. As can be seen above, when using the video recommendation model to recommend videos to users, videos with more interaction data are likely to be recommended to users, while cold start videos in each video are not easily recommended to users due to less interaction data, resulting in a low exposure rate of cold start videos.

[0003] Therefore, there is an urgent need for a video recommendation method to improve the exposure rate of cold start videos. Summary of the Invention

[0004] The purpose of the embodiments of the present application is to provide a video recommendation method, apparatus, electronic device, and storage medium to achieve an improvement in the exposure rate of cold start videos. The specific technical solutions are as follows:

[0005] In the first aspect implemented in the present application, first, a video recommendation method is provided. The method includes:

[0006] In response to a search term of a target user, perform video retrieval based on the search term to obtain multiple videos;

[0007] Select videos with an online duration not greater than a predetermined duration from the multiple videos as cold start videos;

[0008] In response to at least one of the cold start videos, calculate an interest expectation value of the target user for each cold start video based on the user characteristics of the target user and the description characteristics of each cold start; wherein, the description characteristics of each cold start video are the characteristics used to describe the video content of the cold start video, and the interest expectation value is a value used to characterize the degree of interest of the target user;

[0009] Based on the interest expectation value of each cold start video, screen the at least one cold start video to obtain at least one first type of candidate video;

[0010] Based on the at least one first type of candidate video, recommend at least one target recommended video to the target user; wherein, the at least one target recommended video includes at least one first type of candidate video.

[0011] Optionally, recommending at least one target recommended video to the target user based on the at least one first type of candidate video includes:

[0012] Recommending each target recommended video to the target user based on the at least one first type of candidate video and other videos among the multiple videos except the identified cold start video;

[0013] Wherein, each target recommended video includes at least one first type of candidate video and at least one other video.

[0014] Optionally, recommending each target recommended video to the target user based on the at least one first type of candidate video and other videos among the multiple videos except the identified cold start video includes:

[0015] Selecting each second type of candidate video for use as a recommended video from other videos among the multiple videos except the identified cold start video;

[0016] For each video among the selected first type of candidate videos and each second type of candidate video, scoring the video to obtain the target score of the video;

[0017] Based on the obtained target scores of each video, selecting each target recommended video for the target user from the at least one first type of candidate video and each second type of candidate video, and recommending the target recommended video to the target user.

[0018] Optionally, scoring the video to obtain the target score of the video includes:

[0019] Scoring the video from at least two scoring dimensions to obtain the score of each scoring dimension;

[0020] Calculating the target score of the video according to the weight and score of each scoring dimension;

[0021] Wherein, the at least two scoring dimensions include at least two of the degree of relevance of the video to the search term, the upload time of the video, and the score of the video by the video recommendation model.

[0022] Optionally, selecting each target recommended video for the target user from the at least one first type of candidate video and each second type of candidate video based on the obtained target scores of each video includes:

[0023] Sorting the selected first type of candidate videos and each second type of candidate video in descending order of the obtained target scores of each video to obtain a sorted queue;

[0024] When at least one candidate video of the first type is included in the first N videos of the sorting queue, select the first N videos to obtain each target recommended video for the target user;

[0025] When no candidate video of the first type is included in the first N videos of the sorting queue, select the first M1 videos of the sorting queue and select M2 candidate videos of the first type from the sorting queue, and determine the selected M1 videos and M2 candidate videos as each target recommended video for the target user; where the sum of M1 and M2 is N, and the target scores of the selected M2 candidate videos of the first type are greater than the target scores of each candidate video of the first type other than the selected M2 candidate videos of the first type.

[0026] Optionally, the construction method of the user characteristics of the target user includes:

[0027] Determine a plurality of auxiliary videos; where the plurality of auxiliary videos are videos that have been performed any specified action by the target user within a specified time period, and the specified time period is a time period with a predetermined time point as the end time and a predetermined duration;

[0028] Generate the user characteristics of the target user based on the description characteristics of each auxiliary video and the weight coefficients corresponding to the description characteristics of each cold start video; where the weight coefficient corresponding to the description characteristics of each auxiliary video is determined based on the historical behavior of the target user for this auxiliary video.

[0029] Optionally, calculating the interest expectation value of the target user for each cold start video based on the user characteristics of the target user and the description characteristics of each cold start video includes:

[0030] For each cold start video, splice the user characteristics of the target user and the description characteristics of this cold start video to obtain the target joint characteristics of this cold start video;

[0031] Input the target joint characteristics of this cold start video into the interest expectation value evaluation formula to obtain the interest expectation value of the target user for this cold start video.

[0032] Optionally, the interest expectation value evaluation formula is:

[0033]

[0034] where x a represents the target joint characteristics of cold start video a, θ represents the linear regression coefficient, A represents the target covariance matrix, and α represents the balance parameter.

[0035] Optionally, after recommending each target recommended video to the target user, the method further includes:

[0036] Updating the parameters of the interest expectation evaluation formula based on the interaction behavior of the target user with respect to the recommended target recommended videos.

[0037] Optionally, θ = A -1 b; where b is a compensation vector;

[0038] Updating the parameters of the interest expectation evaluation formula based on the interaction behavior of the target user with respect to the recommended target recommended videos includes:

[0039] Adding the A to a first adjustment value to obtain an updated A;

[0040] Adding the b to a second adjustment value to obtain an updated b;

[0041] Where the first adjustment value is: The second adjustment value is rx i ;

[0042] x i is the target joint feature of the target user with respect to any video i belonging to the first type of candidate videos among the target recommended videos recommended for the target user; r is a reward value, and the determination method of r includes: if the target user performs any specified interaction behavior with respect to the video i, then r is 1, otherwise, r is 0.

[0043] The video i is any video belonging to the first type of candidate videos, that is, cold start videos, among the target recommended videos recommended to the target user.

[0044] In a second aspect of the implementation of the present application, there is also provided a video recommendation device, and the device includes:

[0045] A retrieval module, configured to respond to a search term of a target user, and perform video retrieval based on the search term to obtain a plurality of videos;

[0046] A selection module, configured to select, from the plurality of videos, videos with an online duration not greater than a predetermined duration as cold start videos;

[0047] A calculation module, configured to respond to at least one of the cold start videos, and calculate an interest expectation value of the target user for each cold start video based on the user characteristics of the target user and the description features of each cold start video; wherein, the description feature of each cold start video is a feature used to describe the video content of the cold start video, and the interest expectation value is used to characterize the degree of interest of the target user.

[0048] A screening module, configured to screen the at least one cold start video based on the interest expectation value of each cold start video, so as to obtain at least one first - type candidate video;

[0049] A recommendation module, configured to recommend at least one target recommended video to the target user based on the at least one first - type candidate video; wherein, the at least one target recommended video includes at least one first - type candidate video. In a third aspect of the implementation of this application, an electronic device is further provided, including a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete communication with each other through the communication bus;

[0050] The memory is used to store a computer program;

[0051] The processor is configured to implement the video recommendation method described in any one of the above when executing the program stored on the memory.

[0052] In a fourth aspect of the implementation of this application, a computer - readable storage medium is further provided. The computer - readable storage medium stores a computer program, and when the computer program is executed by a processor, the video recommendation method described in any one of the above is implemented.

[0053] In yet another aspect of the implementation of this application, a computer program product including instructions is further provided. When it runs on a computer, it enables the computer to execute the video recommendation method described in any one of the above.

[0054] The solution of this application can respond to the search term of the target user, perform video retrieval based on the search term to obtain multiple videos; select videos with an online duration not greater than a predetermined duration from the multiple videos as cold start videos; in response to at least one of the cold start videos, calculate the interest expectation value of the target user for each cold start video based on the user characteristics of the target user and the description characteristics of each cold start video; screen the at least one cold start video based on the interest expectation value of each cold start video to obtain at least one first - type candidate video; and recommend at least one target recommended video to the target user based on the at least one first - type candidate video. It can be seen that the target recommended videos recommended by the solution of this application to the target user include the cold start videos selected based on the interest expectation value. Therefore, the solution of this application can effectively improve the exposure rate of cold start videos when recommending videos to the target user. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art.

[0056] Figure 1 It is a schematic flowchart of a video recommendation method provided in an embodiment of the present application;

[0057] Figure 2 It is a schematic flowchart of another video recommendation method provided in an embodiment of the present application;

[0058] Figure 3 It is a schematic structural diagram of a video recommendation device provided in an embodiment of the present application;

[0059] Figure 4 It is a schematic structural diagram of an electronic device provided in an embodiment of the present application. Detailed implementation manners

[0060] Next, the technical solutions in the embodiments of the present application will be described in conjunction with the accompanying drawings in the embodiments of the present application.

[0061] In the technical solutions of the present application, the collection, storage, use, processing, transmission, provision, and disclosure of the user's personal information and other processes all comply with the provisions of relevant laws and regulations and do not violate public order and good customs.

[0062] In order to balance the improvement of the exposure rate of cold-start videos and ensure the personalized experience of users, the embodiments of the present application provide a video recommendation method, device, electronic device, and storage medium.

[0063] Next, a video recommendation method provided in an embodiment of the present application will be introduced first. The video recommendation method provided in the embodiment of the present application can be applied to a server. Exemplarily, the server can be the server of a video APP (Application), or the server of a video website. The present application does not limit the server, and the server of the present application can be any server capable of executing the video recommendation method of the present application.

[0064] Among them, the video recommendation method provided in the embodiment of the present application may include:

[0065] In response to the search term of the target user, perform video retrieval based on the search term to obtain multiple videos;

[0066] Select videos with an online duration not greater than a predetermined duration from the multiple videos as cold-start videos;

[0067] In response to at least one of the cold-start videos, calculate the interest expectation value of the target user for each cold-start video based on the user characteristics of the target user and the description characteristics of each cold-start video; where the description characteristics of each cold-start video are the characteristics used to describe the video content of the cold-start video, and the interest expectation value is used to characterize the degree of the target user's interest;

[0068] Screen the at least one cold start video based on the interest expectation value of each cold start video to obtain at least one first - type candidate video;

[0069] Based on the at least one first - type candidate video, recommend at least one target recommended video to the target user; wherein, the at least one target recommended video includes at least one first - type candidate video. The solution of this application can, in response to the search term of the target user, perform video retrieval based on the search term to obtain multiple videos; select videos with an online duration not greater than a predetermined duration from the multiple videos as cold start videos; in response to at least one of the cold start videos, calculate the interest expectation value of the target user for each cold start video based on the user characteristics of the target user and the description characteristics of each cold start video; screen the at least one cold start video based on the interest expectation value of each cold start video to obtain at least one first - type candidate video; and based on the at least one first - type candidate video, recommend at least one target recommended video to the target user. It can be seen that the target recommended videos recommended by the solution of this application to the target user include the cold start videos selected based on the interest expectation value. Therefore, the solution of this application can effectively improve the exposure rate of cold start videos when recommending videos to the target user. As Figure 1 shown, a video recommendation method provided by an embodiment of this application includes:

[0070] S101, in response to the search term of the target user, perform video retrieval based on the search term to obtain multiple videos.

[0071] The application scenario of the solution of this application can be a scenario where a user searches for videos through a search term on any video platform (such as a video APP, a video website, etc.). The server is the server of the video platform (for example, when the video platform is a video APP, the server is the server of this video APP).

[0072] The server can perform word segmentation on the title of each video in the video database, so as to obtain a link for characterizing the corresponding relationship between the word segmentation and the video. Each video in the video database can be uploaded by users and / or staff. For example, if the title of a video is: xxx Beijing Concert, it can be segmented into xxx, Beijing, Concert, and Beijing Concert, etc. Thus, multiple links such as xxx-xxx Beijing Concert, Beijing-xxx Beijing Concert, Concert-xxx Beijing Concert, and Beijing Concert-xxx Beijing Concert can be obtained. If the user's search term is Beijing, then since there is a link of Beijing-xxx Beijing Concert, the video with the title of xxx Beijing Concert can be used as the retrieved video. It can be understood that after performing word segmentation on the titles of multiple videos, there may be the same words, so the same word can correspond to multiple links. For example, for video 1 with the title of xxx Beijing Concert, the link: Concert-xxx Beijing Concert can be obtained, and for video 2 with the title of xx Concert, the link: Concert-xx Concert can be obtained. If the user's search term is Concert, then both video 1 and video 2 can be used as the retrieved videos.

[0073] This application can perform video retrieval based on the preset link between word segmentation and video according to the search term to retrieve multiple videos. Specifically, based on the link, the videos characterized by the link to which each word segmentation that matches the search term belongs can be used as the retrieved multiple videos.

[0074] Specifically, the search term can also be segmented. Exemplarily, if the search term is: Beijing Concert, the search term can be segmented into: Beijing, Concert, and Beijing Concert, etc. Thus, each word segmentation of the search term can be matched with the word segmentations of each video title, and then based on the link, the videos characterized by the link to which the matched word segmentation belongs can be used as the retrieved videos.

[0075] In this application, when performing video retrieval based on the search term, a video retrieval model can also be used for retrieval. The video retrieval model is used to retrieve videos related to the search term according to the search term input by the user.

[0076] Any method that can perform video retrieval based on the search term can be applied to this application, and this application does not limit the method of performing video retrieval based on the search term.

[0077] S102. Select videos with an online duration not greater than a predetermined duration from the multiple videos as cold start videos.

[0078] The cold start video is a video whose online duration is less than a predetermined duration. Videos for cold start can be screened by the video release time. For example, set the screening time window to within half a year, and videos whose release time does not exceed half a year are used as cold start videos. The screening time window and the predetermined duration can be set according to specific scenarios, and the present application does not limit this.

[0079] Furthermore, a video whose online duration is less than a predetermined duration and whose interaction data volume is less than a predetermined quantity can be used as a cold start video. It can be understood that generally, for a cold start video, since its online duration is less than the predetermined duration, the interaction data volume is usually less than the predetermined quantity. However, there are also cases where due to the influence of activities, recommendation strategies, etc. set by the video platform, some videos with an online duration less than the predetermined duration have an interaction data volume not less than the predetermined quantity. Although these videos have an online duration less than the predetermined duration, due to their relatively large interaction data volume, they are no longer cold start videos.

[0080] S103, in response to at least one of the cold start videos, calculate the expected interest value of the target user for each cold start video based on the user characteristics of the target user and the description characteristics of each cold start video; wherein, the description characteristics of each cold start video are the characteristics used to describe the video content of the cold start video, and the expected interest value is a value used to characterize the degree of interest of the target user.

[0081] The at least one cold start video is at least one cold start video selected from the multiple videos.

[0082] In the present application, the user characteristics of the target user are determined based on multiple auxiliary videos historically viewed by the target user.

[0083] Specifically, the construction method of the user characteristics of the target user includes:

[0084] Determine multiple auxiliary videos; wherein, the multiple auxiliary videos are videos for which the target user has performed any specified behavior within a specified time period, and the specified time period is a time period with a predetermined time point as the end time and a predetermined duration.

[0085] Generate the user characteristics of the target user based on the description characteristics of each auxiliary video and the weight coefficient corresponding to the description characteristics of each auxiliary video; wherein, the weight coefficient corresponding to the description characteristics of each auxiliary video is determined based on the historical behavior of the target user for the auxiliary video.

[0086] The specified behavior can be behaviors such as viewing, liking, collecting, forwarding, and reserving.

[0087] In one implementation, the predetermined time point can be used to distinguish expired videos. Videos before the predetermined time point can be expired videos. For example, if the time point is 12:00, then a video at 11:00 before 12:00 is an expired video. The predetermined duration can be set according to the application scenario. Exemplarily, the predetermined duration can be one week. If the predetermined time point is time point A, then the specified time period is the time period from one week after time point A to time point A. At this time, the multiple auxiliary videos are videos on which the target user has performed any specified action within the specified time period. Since the user's interest in videos may change over time, in this embodiment, the predetermined time point is used to distinguish expired videos, so that the selected auxiliary videos can be more matched with the current interest of the target user.

[0088] In another implementation, the multiple auxiliary videos can be sorted according to the time when the target user performs any specified action on the auxiliary, and the Top-N latest videos are selected as the multiple auxiliary videos.

[0089] The weight coefficient of each auxiliary video is determined according to the historical behavior of the target user on the auxiliary video. For example, corresponding coefficient values can be assigned to behaviors such as viewing, liking, favoriting, and sharing. Using data such as the number of views, likes, favorites, and shares of the target user on the auxiliary video, weighted calculation is performed to obtain the weight coefficient of each auxiliary video.

[0090] The description feature of each video is determined in advance based on the content of each video.

[0091] Thus, using the description feature of each auxiliary video and the weight coefficient corresponding to the description feature of each auxiliary video, the user feature of the target user can be generated.

[0092] Optionally, the user feature can be represented by a one-hot vector. For example, corresponding to a 5-dimensional basic description feature vector [a, b, c, d, e], the user's actions on the auxiliary videos are: viewing auxiliary video A, the label of video A includes a, the weight coefficient of video A is alpha, liking auxiliary video B, the label of video B includes b, the weight coefficient of video B is beta, then the user feature vector can be represented as [alpha*1, beta*1, 0, 0, 0]. If the user also views auxiliary video C, the label of video C includes b, and the weight coefficient of video C is omega, then the user feature vector can be represented as [alpha*1, (beta + omega)*1, 0, 0, 0]. The basic feature description vector is determined in advance according to the specific application scenario, and its dimension is not limited in this application.

[0093] Similarly, the vector form corresponding to the description features of the cold start video can also be represented by a one-hot vector. For example, there is a 5-dimensional basic description feature vector [a, b, c, d, e] corresponding to the A video label. The A video description feature vector form can be: [1, 0, 0, 0, 0], and the B video label includes b. The video description feature vector form can be: [0, 1, 0, 0, 0]. The vector corresponding to the description features of the auxiliary video can be predetermined according to the specific application scenario, and its dimension is not limited in this application. It can be understood that in this application, a one-dimensional label is equivalent to descriptive data for one dimension of the video.

[0094] In the present application, the descriptive features of each cold start video are features used to describe the video content of the cold start video, and the descriptive features of each cold start video can be determined by the video tag. Exemplarily, the tags of the cold start video can include: video type (such as suspense, fairy tales, and martial arts, etc.), video length, starring name, director name, etc.

[0095] In this application, whether it is an auxiliary video or a cold start video, the label of each video can be set by the staff according to the video content, or it can be obtained by identifying the video content using a video recognition algorithm. This application does not limit this.

[0096] For each cold start video, this embodiment calculates the expected value of the target user's interest in the cold start video, which is equivalent to predicting the degree of interest of the target user in the cold start video.

[0097] S104: Screen the at least one cold start video based on the expected interest value of each cold start video to obtain at least one first category candidate video.

[0098] There are multiple ways to screen the at least one cold start video. In one way, cold start videos with expected interest values greater than a specified value can be selected as first category candidate videos. In another way, a specified number of cold start videos with the largest expected interest values can be selected as first category candidate videos.

[0099] It can be seen that in this application, the first category of candidate videos screened out for the cold start video using the expected interest value of each cold start video is the video determined in the cold start video to match the interest of the target user, thereby ensuring the user's personalized experience.

[0100] S105, based on the at least one first-category candidate video, recommend at least one target recommended video to the target user; wherein the at least one target recommended video includes at least one first-category candidate video.

[0101] At least one target recommended video recommended to the target user includes at least one cold start video selected according to the interest expectation value of the cold start video.

[0102] The solution of this application can, in response to the search term of the target user, perform video retrieval based on the search term to obtain multiple videos; select, from the multiple videos, videos with an online duration not greater than a predetermined duration as cold start videos; in response to at least one of the cold start videos, calculate the interest expectation value of the target user for each cold start video based on the user characteristics of the target user and the description characteristics of each cold start video; screen the at least one cold start video based on the interest expectation value of each cold start video to obtain at least one first type of candidate video; and recommend at least one target recommended video to the target user based on the at least one first type of candidate video. It can be seen that the target recommended videos recommended to the target user by the solution of this application include the cold start videos selected based on the interest expectation value. Therefore, the solution of this application can effectively improve the exposure rate of the cold start videos when recommending videos to the target user.

[0103] In addition, since the cold start videos in the target recommended videos recommended to the target user are the determined cold start videos that match the interests of the target user, the solution of this application can ensure the personalized experience of the user.

[0104] Optionally, there can be various ways to calculate the interest expectation value of the target user for each cold start video based on the user characteristics of the target user and the description characteristics of each cold start video.

[0105] In one implementation, calculating the interest expectation value of the target user for each cold start video based on the user characteristics of the target user and the description characteristics of each cold start video includes:

[0106] For each cold start video, splice the user characteristics of the target user and the description characteristics of this cold start video to obtain the target joint characteristics of this cold start video;

[0107] Input the target joint characteristics of this cold start video into the interest expectation value evaluation formula to obtain the interest expectation value of the target user for this cold start video.

[0108] The user characteristics can be used to represent the user's interest tendency for videos, and the description characteristics of each cold start video can be used to describe this cold start video. Therefore, the user characteristics and the description characteristics of the cold start video can be used to predict the interest expectation value of the target user for this cold start video.

[0109] In this implementation, the LinUCB algorithm (the LinUCB algorithm is a context-aware multi-armed bandit algorithm. The LinUCB algorithm can dynamically select the videos to be recommended according to the user's characteristics and the characteristics of the videos, and can balance exploration and exploitation, so as to achieve personalized recommendation) can be used to estimate the expected revenue of each cold-start video shown to the user according to the user's descriptive characteristics, the descriptive characteristics of the videos, and the user's behavior, so as to select the optimal new video for distribution. The core idea of this algorithm is to use linear regression to estimate the expected revenue of each new video based on the context, and adjust the balance between exploration and exploitation according to the confidence level. Specifically, based on the concept of this algorithm, the interest expectation value evaluation formula can be used to calculate the interest expectation value.

[0110] Specifically, the interest expectation value evaluation formula is as follows:

[0111]

[0112] where x a represents the target joint feature of the cold-start video a, θ represents the linear regression coefficient, A represents the target covariance matrix, and α represents the balance parameter. is the transpose of x a and A -1 is the inverse of A. x a and θ can be in vector form.

[0113] can be called the exploitation score, which can be used to directly estimate the degree of interest of the user in this video depending on the current model parameters and user characteristics.

[0114] can be called the exploration score, which can increase the additional exploration opportunities for cold-start videos with higher uncertainty based on the uncertainty of the current model to avoid the model converging to the local optimal solution prematurely.

[0115] In this application, the interest expectation value evaluation formula can update the parameters of the formula based on the interaction behavior of the target user for the recommended target recommended video. Therefore, optionally, after recommending each target recommended video to the target user, the method further includes:

[0116] Updating the parameters of the interest expectation value evaluation formula based on the interaction behavior of the target user for the recommended target recommended video.

[0117] Specifically, θ = A -1 b; where b is the compensation vector;

[0118] Updating the parameters of the interest expectation value evaluation formula based on the interaction behavior of the target user with respect to the recommended target recommended video, including:

[0119] Adding the A to the first adjustment value to obtain the updated A;

[0120] Adding the b to the second adjustment value to obtain the updated b;

[0121] Wherein, the first adjustment value is: The second adjustment value is: rx i ;

[0122] x i is the target joint feature of any video i belonging to the first type of candidate videos in the target recommended video recommended for the target user; r is the reward value, and the determination method of r includes: if the target user performs any specified interaction behavior for the video i, then r is 1, otherwise, r is 0.

[0123] The initial value of the A can be the identity matrix, and the initial value of the b can be the all-zero vector.

[0124] It should be emphasized that for each target user, there can be one interest expectation value evaluation formula corresponding to each user, or all users can share one interest expectation value evaluation formula, that is, the interaction behavior performed by each user with respect to the target recommended video recommended to the user can affect the parameters of the interest expectation value evaluation formula.

[0125] In one implementation manner, calculating the interest expectation value of the target user for each cold start video based on the user feature of the target user and the description feature of each cold start video includes:

[0126] Inputting the user feature of the target user and the description feature of each cold start video into the interest expectation value prediction model to obtain the interest expectation value of the target user for each cold start video.

[0127] The interest expectation value prediction model is trained using multiple sample features and the true value of each sample feature. Among them, the sample feature is a feature obtained by splicing the feature of the sample user and the feature of the sample cold start video, and the true value of each sample feature is the true interest expectation value of the sample user corresponding to the sample feature for the sample cold start video corresponding to the sample feature.

[0128] It can be seen that the solution of this embodiment can effectively determine, for each cold start video, the interest expectation value representing the degree of interest of the target user in the cold start video.

[0129] Optionally, there are various implementation manners for recommending at least one target recommended video to the target user based on the at least one first type of candidate video.

[0130] In one implementation manner, at least one target recommended video recommended to the target user may only include the at least one first type of candidate video, that is, only include cold start videos.

[0131] In one implementation manner, the recommending at least one target recommended video to the target user based on the at least one first type of candidate video includes:

[0132] Recommending each target recommended video to the target user based on the at least one first type of candidate video and other videos in the multiple videos except the identified cold start videos;

[0133] Wherein, each target recommended video includes at least one first type of candidate video and at least one other video.

[0134] In this embodiment, among the target recommended videos recommended to the user, there are not only cold start videos but also non-cold start videos.

[0135] Specifically, the recommending each target recommended video to the target user based on the at least one first type of candidate video and other videos in the multiple videos except the identified cold start videos includes steps A1 - A3.

[0136] Step A1, select each second type of candidate video used as a recommended video from other videos in the multiple videos except the identified cold start videos;

[0137] The other videos are non-cold start videos. There are various manners for selecting non-cold start videos. In one implementation manner, it can be selected in the same manner as the selection manner for cold start videos above. For each non-cold start video, the expected interest value can be calculated, and each second type of candidate video used as a recommended video can be selected according to the expected interest value.

[0138] Another manner for selecting non-cold start videos is to use a video recommendation model to select each second type of candidate video used as a recommended video.

[0139] In this application, although the video recommendation model has relatively weak recommendation ability for cold start videos, it is relatively accurate for recommending non-cold start videos. Therefore, in this application, for non-cold start videos, the video recommendation model is used for analysis to obtain each second type of candidate video used as a recommended video.

[0140] Generally, a video recommendation model can score the videos to be analyzed. The higher the score of a video, the more it represents a match with the target user. It can be understood that the input of the video recommendation model may include target user features and video features of the videos to be analyzed. The target user features may include the user's age, gender, historical viewing behavior, viewing interests, etc. The video features may include: description features of the video, interaction data of the video, etc. And precisely because the non-cold start videos have richer dimensions and more data volume of video features compared to cold start videos, the video recommendation model is more accurate in recommending non-cold start videos.

[0141] The video recommendation model itself is not an improvement point of this application. Any video recommendation model that scores videos based on interaction data such as the click-through rate, view count, and like count of the video can be applied to the solution of this application.

[0142] Step A2, for each of the selected first-class candidate videos and each of the second-class candidate videos, score the video to obtain the target score of the video.

[0143] Specifically, the scoring of the video to obtain the target score of the video includes:

[0144] Score the video from at least two scoring dimensions to obtain the score of each scoring dimension;

[0145] Calculate the target score of the video according to the weight and score of each scoring dimension;

[0146] Among them, the at least two scoring dimensions include at least two of the degree of relevance of the video to the search term, the upload time of the video, and the score given by the video recommendation model to the video.

[0147] For each video, the calculation method of the score of the scoring dimension of the degree of relevance of the video to the search term may include:

[0148] Calculate the vector similarity between the title of the video and the search term to obtain the score representing the degree of relevance of the video to the search term.

[0149] For each video, the calculation method of the score of the upload time dimension of the video may include:

[0150] Assign a score according to the upload time of the video. The later the upload time of the video, the higher the assigned score.

[0151] The score given by the video recommendation model to the video has been introduced in the above embodiments and will not be elaborated here.

[0152] It can be understood that, since the video interaction data in the first type of candidate videos is relatively small, the scores given by the video recommendation model to the videos in the first type of candidate videos are lower than those of the videos in the second type of candidate videos.

[0153] Therefore, the present application can set the weight of the score for each dimension according to the specific application scenario to set the influence degree of each dimension on the target score.

[0154] Exemplarily, the target score of a video is calculated using three scoring dimensions: the relevance degree of the video to the search term, the upload time of the video, and the score given by the video recommendation model to the video. The score of the relevance degree of the video to the search term dimension is a, the score of the upload time dimension of the video is b, the score given by the video recommendation model to the video is c, the weight of the relevance degree of the video to the search term dimension is 3, the weight of the upload time dimension of the video is 1, and the weight of the score given by the video recommendation model to the video is 2. Then the target score is: 3×a + 1×b + 2×c.

[0155] After obtaining the target scores of each video, the target recommended videos to be recommended to the target user can be determined according to the target scores of each video.

[0156] Step A3, based on the obtained target scores of each video, select the target recommended videos for the target user from the at least one first type of candidate videos and each second type of candidate videos, and recommend the target recommended videos to the target user.

[0157] The selecting the target recommended videos for the target user from the at least one first type of candidate videos and each second type of candidate videos based on the obtained target scores of each video includes:

[0158] Sort the selected first type of candidate videos and each second type of candidate videos in descending order of the target scores of the obtained videos to obtain a sorted queue;

[0159] In the case that at least one first type of candidate video is included in the first N videos in the sorted queue, select the first N videos to obtain the target recommended videos for the target user;

[0160] In the case that none of the first - type candidate videos is included in the first N videos of the sorting queue, select the first M1 videos of the sorting queue and select M2 first - type candidate videos from the sorting queue, and determine the selected M1 videos and M2 first - type candidate videos as the respective target recommended videos for the target user; where the sum of M1 and M2 is N, and the target scores of the selected M2 first - type candidate videos are greater than the target scores of each of the first - type candidate videos other than the selected M2 first - type candidate videos among the selected first - type candidate videos.

[0161] It can be understood that although the recommendation model has relatively weak recommendation ability for cold - start videos, it is relatively accurate for non - cold - start videos. Therefore, when determining the recommended videos, considering the matching degree between the target user and the recommended videos, the two scoring dimensions of the relevance degree between the video and the search term and the score of the video by the video recommendation model have a greater impact when calculating the target score. Therefore, the target scores of the second - type candidate videos are generally relatively high compared to the first - type candidate videos. There may be a situation where none of the first - type candidate videos is included in the first N videos of the sorting queue. At this time, the first M1 videos of the sorting queue can be selected, and M2 first - type candidate videos can be selected from the sorting queue, and the selected M1 videos and M2 first - type candidate videos are determined as the respective target recommended videos for the target user; where the sum of M1 and M2 is N.

[0162] It can be understood that the first M1 videos all belong to the second - type candidate videos, and the selected M2 first - type candidate videos are the M2 first - type candidate videos with the highest comprehensive scores among the first - type candidate videos.

[0163] Each of the target recommended videos recommended to the target user can be arranged in the order corresponding to the descending order of the target scores of each target recommended video and displayed on the client of the target user. Thus, the target user can first see the video with the highest comprehensive score, that is, the video with a relatively high matching degree with the target user, which can improve the usage experience of the target user. For the first - type candidate videos recommended to the target user, that is, cold - start videos, if the target user is interested in them, corresponding specified behaviors will be executed, and thus, the behavioral data of the target user for this video can be increased.

[0164] For example: set N = 8, is the second type of candidate video is the first type of candidate video. After arranging through the comprehensive scoring mechanism and using the target score, the video list becomes , since the display position can expose at most N = 8 videos, Shown to the target user. Thus, the solution of this embodiment can dynamically adjust the display position of the recommended videos through a comprehensive scoring mechanism.

[0165] In the solution of this embodiment, the target recommended videos recommended to the target user include cold start videos and non-cold start videos that match the personalization of the target user. Moreover, the recommended cold start videos are also videos that match the interests of the target user determined according to the interest expectation value. Thus, the solution of this embodiment can take into account improving the exposure rate of cold start videos and ensuring the personalized experience of users when recommending videos to the target user.

[0166] Next, through a specific embodiment, a specific introduction to the video recommendation method will be given. Figure 2 It is a schematic flowchart of another video recommendation method provided in the embodiment of the present application. As Figure 2 shown, a video recommendation method provided in the embodiment of the present application may include the following steps:

[0167] S201. Receive the search term input by the target user;

[0168] S202. Segment the search term;

[0169] S203. Key-value recall;

[0170] Steps S201 - S202 correspond to S101 above and are used to respond to the search term of the target user and perform video retrieval based on the search term to obtain multiple videos. Details are not elaborated here.

[0171] S204. Screen videos;

[0172] Corresponds to S102 and step A1. For the videos determined to match the search term, the first type of candidate videos belonging to cold start videos and the second type of candidate videos belonging to non - cold start videos can be screened out.

[0173] For cold start videos, execute S205 - S207; for non - cold start videos, execute S208 - S209.

[0174] S205. Generate a user profile and a new video feature vector;

[0175] Corresponds to determining the user characteristics of the target user and the description characteristics of cold start videos above. Details are not elaborated here.

[0176] S206. Score using a scoring algorithm;

[0177] S206 corresponds to calculating the expected interest value of the target user for each cold start video above. The scoring refers to determining the expected interest value of each cold start video. The scoring algorithm can be the LinUCB algorithm. The LinUCB algorithm was introduced in the above embodiments and details are not elaborated here.

[0178] S207. Retain the M videos with the highest scores;

[0179] Corresponds to step S104. Specifically, the M videos with the highest scores can be retained as the first type of candidate videos.

[0180] S208. Score using the ranking model;

[0181] S209. Retain the N videos with the highest scores;

[0182] S208 - S209 correspond to step A1 above. The ranking model is the video recommendation model. In this step, the ranking model is used to select the N videos with the highest scores from non - cold start videos as the second type of candidate videos.

[0183] S210, local reordering;

[0184] S211, presenting the video to the target user.

[0185] S210 - S211 corresponds to the above-mentioned steps A2 - A3, and will not be elaborated here.

[0186] The solution of this embodiment deeply integrates the user's historical behavior into cold start, establishes a multi-dimensional matching model between the user and the video content, and efficiently tracks the evolution of the user's interests; it can break through the limitation of fixed-position recommendation for video distribution and achieve dynamic and intelligent distribution of cold-start videos. The solution of this application does not rely on complex models, so it can quickly respond to the changes in the user's interests and significantly improve the real-time performance of recommendations.

[0187] Based on the content of the embodiment of the above video recommendation method, an embodiment of this application also provides a video recommendation device. Figure 3 As shown in the structure schematic diagram of a video recommendation device provided by an embodiment of this application, Figure 3 as shown, the video recommendation device may include:

[0188] A retrieval module 301, configured to respond to a search term of a target user, perform video retrieval based on the search term to obtain multiple videos;

[0189] A selection module 302, configured to select, from the multiple videos, videos with an online duration not greater than a predetermined duration as cold-start videos;

[0190] A calculation module 303, configured to respond to at least one of the cold-start videos, calculate an interest expectation value of the target user for each cold-start video based on the user characteristics of the target user and the description characteristics of each cold-start video; wherein, the description characteristics of each cold-start video are the characteristics used to describe the video content of the cold-start video, and the interest expectation value is used to characterize the degree of the target user's interest;

[0191] A screening module 304, configured to screen the at least one cold-start video based on the interest expectation value of each cold-start video to obtain at least one first-class candidate video;

[0192] A recommendation module 305, configured to recommend at least one target recommended video to the target user based on the at least one first-class candidate video; wherein, the at least one target recommended video includes at least one first-class candidate video.

[0193] The solution of this application can, in response to the search term of the target user, perform video retrieval based on the search term to obtain multiple videos; select videos with an online duration not greater than a predetermined duration from the multiple videos as cold start videos; respond to at least one of the cold start videos, and calculate the expected interest value of the target user for each cold start video based on the user characteristics of the target user and the description characteristics of each cold start video; screen the at least one cold start video based on the expected interest value of each cold start video to obtain at least one first type of candidate video; and recommend at least one target recommended video to the target user based on the at least one first type of candidate video. It can be seen that the target recommended videos recommended by the solution of this application include the cold start videos selected based on the expected interest value. Therefore, the solution of this application can effectively improve the exposure rate of cold start videos when recommending videos to the target user.

[0194] Optionally, the recommendation module includes:

[0195] A recommendation sub-module, configured to recommend each target recommended video to the target user based on the at least one first type of candidate video and other videos in the multiple videos except the identified cold start videos;

[0196] Wherein, each target recommended video includes at least one first type of candidate video and at least one other video.

[0197] Optionally, the recommendation sub-module includes:

[0198] A selection unit, configured to select each second type of candidate video used as a recommended video from other videos in the multiple videos except the identified cold start videos;

[0199] A scoring unit, configured to score each video among the selected first type of candidate videos and each second type of candidate video to obtain the target score of the video;

[0200] A recommendation unit, configured to select each target recommended video for the target user from the at least one first type of candidate video and each second type of candidate video based on the obtained target scores of the videos, and recommend the target recommended videos to the target user.

[0201] Optionally, the scoring unit is specifically configured to:

[0202] For each video among the selected first type of candidate videos and each second type of candidate video, score the video from at least two scoring dimensions to obtain the score of each scoring dimension;

[0203] Calculate the target score of the video according to the weights and scores of each scoring dimension;

[0204] Among them, the at least two scoring dimensions include at least two of the relevance of the video to the search term, the upload time of the video, and the score of the video by the video recommendation model.

[0205] Optionally, the recommendation unit includes:

[0206] A sorting subunit, configured to sort the selected first-class candidate videos and each second-class candidate video in a descending order of the target scores of the obtained videos to obtain a sorted queue;

[0207] A selection subunit, configured to select the first N videos in the sorted queue when at least one first-class candidate video is included in the first N videos in the sorted queue, to obtain each target recommended video for the target user, and recommend the target recommended videos to the target user;

[0208] When no first-class candidate video is included in the first N videos in the sorted queue, select the first M1 videos in the sorted queue, and select M2 first-class candidate videos from the sorted queue, and determine the selected M1 videos and M2 first-class candidate videos as each target recommended video for the target user; where the sum of M1 and M2 is N, and the target scores of the selected M2 first-class candidate videos are greater than the target scores of each first-class candidate video other than the selected M2 first-class candidate videos among the selected first-class candidate videos.

[0209] Optionally, the construction method of the user characteristics of the target user includes:

[0210] Determine multiple auxiliary videos; where the multiple auxiliary videos are videos that have been performed any specified behavior by the target user within a specified time period, and the specified time period is a time period with a predetermined time point as the end time and a predetermined duration;

[0211] Generate the user characteristics of the target user based on the description features of each auxiliary video and the weight coefficients corresponding to the description features of each cold start video; where the weight coefficient corresponding to the description feature of each auxiliary video is determined based on the historical behavior of the target user on the auxiliary video.

[0212] Optionally, the calculation module includes:

[0213] A splicing sub-module, configured to splice the user characteristics of the target user and the description features of the cold start video for each cold start video to obtain the target joint features of the cold start video;

[0214] An input subunit, configured to input the target combined feature of the cold start video into an interest expectation value evaluation formula, and obtain the interest expectation value of the target user for the cold start video.

[0215] Optionally, the interest expectation value evaluation formula is:

[0216]

[0217] where x a represents the target combined feature of the cold start video a, θ represents the linear regression coefficient, A represents the target covariance matrix, and α represents the balance parameter.

[0218] Optionally, the apparatus further includes:

[0219] An update module, configured to update the parameters of the interest expectation value evaluation formula based on the interaction behavior of the target user for the target recommended videos recommended to the target user after recommending each target recommended video to the target user.

[0220] Optionally, the θ = A -1 b; where b is a compensation vector;

[0221] Updating the parameters of the interest expectation value evaluation formula based on the interaction behavior of the target user for the target recommended videos recommended to the target user includes:

[0222] Adding the A and a first adjustment value to obtain an updated A;

[0223] Adding the b and a second adjustment value to obtain an updated b;

[0224] where the first adjustment value is: The second adjustment value is: rx i ;

[0225] x i is the target combined feature of the target user for any video i belonging to the first type of candidate videos among the target recommended videos recommended to the target user; r is a reward value, and the determination method of r includes: if the target user performs any specified interaction behavior for the video i, then r is 1, otherwise, r is 0.

[0226] An embodiment of the present application further provides an electronic device, as Figure 4 shown, including a processor 401, a communication interface 402, a memory 403, and a communication bus 404, where the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404,

[0227] The memory 403 is used to store a computer program;

[0228] The processor 401 is configured to implement the video recommendation method described in any one of the above when executing the program stored in the memory 403.

[0229] The communication bus mentioned in the above terminal may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of simplicity, only a thick line is used to represent it in the figure, but it does not mean that there is only one bus or one type of bus.

[0230] The communication interface is used for communication between the above terminal and other devices.

[0231] The memory may include a Random Access Memory (RAM), or may also include a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located far from the aforementioned processor.

[0232] The above-mentioned processor may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0233] In another embodiment provided by the present application, there is also provided a computer-readable storage medium, in which a computer program is stored, and when the computer program is executed by a processor, it implements the video recommendation method described in any one of the above embodiments.

[0234] In another embodiment provided by the present application, there is also provided a computer program product containing instructions, which when running on a computer, causes the computer to execute the video recommendation method described in any one of the above embodiments.

[0235] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access, or a data storage device such as a server or data center that includes one or more integrated available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)).

[0236] It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including", or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or device that includes a series of elements includes not only those elements but also other elements that are not expressly listed, or also includes elements that are inherent to such process, method, article, or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article, or device that includes the element.

[0237] Each embodiment in this specification is described in a related manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments.

[0238] The above description is only a preferred embodiment of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application are included in the protection scope of the present application.

Claims

1. A video recommendation method, characterized in that, The method includes: In response to a search term of a target user, performing video retrieval based on the search term to obtain multiple videos; Selecting, from the multiple videos, videos with an online duration not greater than a predetermined duration as cold start videos; In response to at least one of the cold start videos, calculating, based on the user characteristics of the target user and the description features of each cold start video, an expected interest value of the target user for each cold start video; wherein, the description feature of each cold start video is a feature for describing the video content of the cold start video, and the expected interest value is a value used to characterize the degree of interest of the target user; Based on the expected interest value of each cold start video, screening the at least one cold start video to obtain at least one first - type candidate video; Based on the at least one first - type candidate video, recommending at least one target recommended video to the target user; wherein, the at least one target recommended video includes at least one first - type candidate video.

2. The method according to claim 1, wherein The recommending at least one target recommended video to the target user based on the at least one first - type candidate video includes: Based on the at least one first - type candidate video and other videos in the multiple videos except the identified cold start videos, recommending each target recommended video to the target user; Wherein, each target recommended video includes at least one first - type candidate video and at least one other video.

3. The method according to claim 2, characterized in that Based on the at least one first - type candidate video and other videos in the multiple videos except the identified cold start videos, recommending each target recommended video to the target user, includes: Selecting, from other videos in the multiple videos except the identified cold start videos, each second - type candidate video to be used as a recommended video; For each video among the selected first - type candidate videos and each second - type candidate video, scoring the video to obtain the target score of the video; Based on the obtained target scores of each video, selecting, from the at least one first - type candidate video and each second - type candidate video, each target recommended video for the target user, and recommending the target recommended video to the target user.

4. The method according to claim 3, wherein The scoring the video to obtain the target score of the video includes: Scoring the video from at least two scoring dimensions to obtain the score of each scoring dimension; Calculating the target score of the video according to the weight and score of each scoring dimension; Wherein, the at least two scoring dimensions include at least two of the relevance degree of the video to the search term, the upload time of the video, and the score of the video by a video recommendation model.

5. The method according to claim 3, characterized in that The selecting, from the at least one first - type candidate video and each second - type candidate video, each target recommended video for the target user based on the obtained target scores of each video includes: Sorting the selected first - type candidate videos and each second - type candidate video in descending order of the target scores of the obtained videos to obtain a sorted queue; When at least one candidate video of the first type is included in the first N videos of the sorting queue, select the first N videos to obtain each target recommended video for the target user; When none of the candidate videos of the first type is included in the first N videos of the sorting queue, select the first M1 videos of the sorting queue and select M2 candidate videos of the first type from the sorting queue, and determine the selected M1 videos and M2 candidate videos of the first type as each target recommended video for the target user; where the sum of M1 and M2 is N, and the target scores of the selected M2 candidate videos of the first type are greater than the target scores of each of the candidate videos of the first type other than the selected M2 candidate videos of the first type.

6. The method according to claim 1, wherein The construction method of the user characteristics of the target user includes: Determine multiple auxiliary videos; where the multiple auxiliary videos are videos that have been performed any specified behavior by the target user within a specified time period, and the specified time period is a time period with a predetermined time point as the end time and a predetermined duration; Generate the user characteristics of the target user based on the description characteristics of each auxiliary video and the weight coefficient corresponding to the description characteristics of each auxiliary video; where the weight coefficient corresponding to the description characteristics of each auxiliary video is determined based on the historical behavior of the target user for this auxiliary video.

7. The method according to claim 1, characterized in that The calculating, based on the user characteristics of the target user and the description characteristics of each cold start video, the interest expectation value of the target user for each target video includes: For each cold start video, splice the user characteristics of the target user and the description characteristics of this target video to obtain the target joint characteristics of this cold start video; Input the target joint characteristics of this cold start video into the interest expectation value evaluation formula to obtain the interest expectation value of the target user for this cold start video.

8. The method according to claim 7, wherein The interest expectation value evaluation formula is: Among them, x a represents the target joint feature of the cold start video a, θ represents the linear regression coefficient, A represents the target covariance matrix, and α represents the balance parameter.

9. The method according to claim 8, wherein After recommending each target recommended video to the target user, the method further includes: Update the parameters of the interest expectation value evaluation formula based on the interaction behavior of the target user for the recommended target recommended videos.

10. The method according to claim 9, characterized in that where θ = A -1 b; where b is a compensation vector; Updating the parameters of the interest expectation value evaluation formula based on the interaction behavior of the target user for the recommended target recommended videos includes: Add the A to the first adjustment value to obtain the updated A; Add the b to the second adjustment value to obtain the updated b; Among them, the first adjustment value is: The second adjustment value is: rx i ; x i is the target combined feature of video i, which belongs to the first type of candidate videos, in the target recommended videos recommended for the target user; r is the reward value, and the determination method of r includes: if the target user performs any specified interaction behavior for this video i, then r is 1, otherwise, r is 0.

11. A video recommendation device, characterized in that, The device includes: A retrieval module, configured to respond to a search term of a target user, perform video retrieval based on the search term to obtain multiple videos; A selection module, configured to select videos with an online duration not greater than a predetermined duration from the multiple videos as cold start videos; A calculation module, configured to respond to at least one of the cold start videos, and calculate an expected interest value of the target user for each cold start video based on the user characteristics of the target user and the description features of each cold start video; wherein, the description feature of each cold start video is a feature used to describe the video content of the cold start video, and the expected interest value is a value used to characterize the degree of interest of the target user. A screening module, configured to screen the at least one cold start video based on the expected interest value of each cold start video to obtain at least one first type of candidate video. A recommendation module, configured to recommend at least one target recommended video to the target user based on the at least one first type of candidate video; wherein, the at least one target recommended video includes at least one first type of candidate video.

12. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete mutual communication through the communication bus. The memory is used to store computer programs. The processor is configured to implement the method according to any one of claims 1-10 when executing the program stored on the memory.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the method according to any one of claims 1-10 is implemented.

Images