Video recommendation method and device

Through the fusion of multiple sorting models and the association of positive and negative training samples, the low accuracy problem caused by the independence of the rough and fine sorting models in the video recommendation system is solved, and the accuracy and user experience of video recommendation are improved.

CN113158020BActive Publication Date: 2024-12-03BEIJING DAJIA INTERNET INFORMATION TECH CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202010074374.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-01-22
Publication Date
2024-12-03
Estimated Expiration
2040-06-03

AI Technical Summary

Technical Problem

In the existing video recommendation system, the training process of the rough-scheduling model and the fine-scheduling model are independent of each other, resulting in the user-recommended target videos with low accuracy, and the rough-scheduling stage may filter out videos with higher ratings in the fine-scheduling stage.

Method used

Using multiple pre-trained sorting models, by inputting user features and video features into sorting models with different precisions, the fusion sorting results are generated to ensure that the rough sorting stage does not filter videos with higher scores, and the second is determined through the third sorting model. Sort the positive and negative training samples of the model to improve recommendation accuracy.

Benefits of technology

It improves the accuracy of video recommendations and ensures that videos with higher ratings in the rough scheduling stage are not missed in the fine scheduling stage, thereby improving the target video quality recommended by users.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113158020B_ABST
    Figure CN113158020B_ABST
Patent Text Reader

Abstract

The present disclosure is about a method and device for recommending videos, which includes: obtaining user features of a target user and video features of each candidate video; inputting the user features of the target user and the video features of the candidate video into a pre-trained first sorting model and a second sorting model, respectively, to obtain a first sorting result and a second sorting result of the candidate video; determining a fusion sorting result of each candidate video according to the first sorting result and the second sorting result of each candidate video, and selecting a first preset number of candidate videos from each candidate video as target videos of the target user; inputting the user features of the target user and the video features of the target video into a third sorting model, to obtain a third sorting result of the target video; and recommending videos to the target user according to the third sorting result of each target video. The present disclosure can improve the accuracy of the target videos recommended for the user.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a video recommendation method and device. Background Art

[0002] At present, the traditional video recommendation process can be divided into the recall stage, the rough ranking stage, the fine ranking stage and the strategy stage. Among them, the recall stage is to screen out the first candidate video that the user may be interested in from the massive videos stored in the video library; the rough ranking stage is to roughly score the first candidate video according to the preset rough ranking model, and screen out the second candidate video from the first candidate video in descending order of scores; the fine ranking stage is to accurately score the second candidate video according to the preset fine ranking model, and screen out the third candidate video from the second candidate video in descending order of scores; the strategy stage is to select the target video from the third candidate video according to the preset recommendation strategy, and recommend the target video to the user.

[0003] However, since the training processes of the coarse ranking model and the fine ranking model are independent of each other and use the same sample data for training, the coarse ranking stage will filter out videos with higher scores in the fine ranking stage from the first candidate videos, resulting in a lower accuracy rate of the target videos recommended to users. Summary of the invention

[0004] The present disclosure provides a video recommendation method and device to at least solve the problem of low accuracy of target videos recommended to users in the related art. The technical solution of the present disclosure is as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, a method for recommending a video is provided, the method comprising:

[0006] Obtaining user features of the target user and video features of each candidate video;

[0007] For each candidate video among the candidate videos, inputting the user features of the target user and the video features of the candidate video into a pre-trained first sorting model and a second sorting model respectively, to obtain a first sorting result and a second sorting result of the candidate video, wherein the positive and negative training samples of the second sorting model are determined according to a pre-trained third sorting model, and the sorting accuracy of the third sorting model is higher than the sorting accuracy of the first sorting model and the second sorting model;

[0008] Determine a fusion ranking result of the candidate videos according to the first ranking result and the second ranking result of the candidate videos, and select a first preset number of candidate videos from the candidate videos in descending order of the fusion ranking results as target videos of the target user;

[0009] For each target video in each target video, input the user feature of the target user and the video feature of the target video into the third sorting model to obtain a third sorting result of the target video;

[0010] Recommending videos to the target user according to the third ranking results of the target videos.

[0011] According to a second aspect of an embodiment of the present disclosure, a video rough sorting device is provided, the device comprising:

[0012] A first acquisition module is configured to acquire user features of a target user and video features of each candidate video;

[0013] A first input module is configured to execute, for each candidate video among the candidate videos, inputting the user features of the target user and the video features of the candidate video into a pre-trained first sorting model and a second sorting model, respectively, to obtain a first sorting result and a second sorting result of the candidate video, wherein the positive and negative training samples of the second sorting model are determined according to a pre-trained third sorting model, and the sorting accuracy of the third sorting model is higher than the sorting accuracy of the first sorting model and the second sorting model;

[0014] A selection module is configured to determine a fusion ranking result of the candidate videos according to the first ranking result and the second ranking result of the candidate videos, and select a first preset number of candidate videos from the candidate videos in descending order of the fusion ranking results as target videos of the target user;

[0015] A second input module is configured to input the user features of the target user and the video features of the target video into the third sorting model for each target video in the target videos, so as to obtain a third sorting result of the target video;

[0016] The recommendation module is configured to recommend videos to the target user according to the third ranking results of the target videos.

[0017] According to a third aspect of an embodiment of the present disclosure, a video server is provided, the video server comprising:

[0018] processor;

[0019] a memory for storing instructions executable by the processor;

[0020] The processor is configured to execute the instructions to implement the rough arrangement method of videos described in any one of the first aspects.

[0021] According to a fourth aspect of an embodiment of the present disclosure, a storage medium is provided. When instructions in the storage medium are executed by a processor of a video server, the video server can execute the video recommendation method described in any one of the first aspects.

[0022] According to a fifth aspect of the embodiments of the present disclosure, a computer program product is provided. When the computer program product is executed on a computer, the computer executes any one of the video recommendation methods described in the first aspect.

[0023] The technical solution provided by the embodiment of the present disclosure brings at least the following beneficial effects: The embodiment of the present disclosure provides a method and device for recommending videos. The video server obtains the user features of the target user and the video features of each candidate video. Then, for each candidate video in the recommended candidate videos of the video, the video server inputs the user features of the recommended target user of the video and the video features of the candidate video into the pre-trained first sorting model and the second sorting model, respectively, to obtain the first sorting result and the second sorting result of the candidate video. Among them, the positive and negative training samples of the recommended second sorting model of the video are determined according to the pre-trained third sorting model, and the sorting accuracy of the recommended third sorting model of the video is higher than the sorting accuracy of the recommended first sorting model of the video and the recommended second sorting model of the video. Afterwards, the video server determines the fusion sorting results of the recommended candidate videos of the video according to the first sorting results and the second sorting results of the recommended candidate videos of the video, and selects the first preset number of candidate videos from the recommended candidate videos of the video in the order of the recommended fusion sorting results from high to low as the target videos of the recommended target user of the video. Finally, for each target video in each target video, the video server inputs the user features of the target user of the video recommendation and the video features of the target video into the third ranking model of the video recommendation, obtains the third ranking result of the target video, and recommends the video to the target user of the video recommendation according to the third ranking result of each target video of the video recommendation. In this way, since the training processes of the second ranking model and the third ranking model are interrelated, and the positive and negative training samples used by the second ranking model are determined by the third ranking model, it is ensured that the rough ranking stage will not filter out the videos with higher scores in the fine ranking stage among the candidate videos, thereby improving the accuracy of the target videos recommended for the user.

[0024] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0026] Figure 1 is a flowchart of a video recommendation method according to an exemplary embodiment;

[0027] Figure 2 is a flowchart of a training method for a second sorting model according to an exemplary embodiment;

[0028] Figure 3 is a block diagram of a video recommendation device according to an exemplary embodiment;

[0029] Figure 4 is a block diagram of a video recommendation device according to an exemplary embodiment;

[0030] Figure 5 The block diagram of a video server is shown according to an exemplary embodiment. DETAILED DESCRIPTION

[0031] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0032] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0033] Figure 1 is a flowchart of a video recommendation method according to an exemplary embodiment. Figure 1 As shown, the method may include the following steps.

[0034] Step 101: Obtain user features of a target user and video features of each candidate video.

[0035] In implementation, when a user (i.e., target user) logs in to the video server through a video client installed on a terminal device, the video server can obtain the user characteristics of the target user. Among them, the user characteristics may include user identification, click-through rate, like rate and attention rate, as well as the identification of the browsed video, the identification of the author, click-through rate, like rate and attention rate. Then, the video server may screen the videos (i.e., candidate videos) that the target user is interested in in the video library according to the user characteristics of the target user, and further obtain the video characteristics of each candidate video. Among them, the video characteristics may include the identification of the video, the identification of the author, click-through rate, like rate and attention rate, as well as the user identification, click-through rate, like rate and attention rate of the user who browses the video. Optionally, the identification of the user, the identification of the browsed video, and the identification of the author of the browsed video may be an embedding vector obtained by the video server through a factor machine (FM) algorithm. In addition, the video server may also screen the candidate videos that the target user is interested in in the video library according to the model of the user terminal of the target user, the network type and other information.

[0036] Step 102: for each candidate video in the candidate videos, the user features of the target user and the video features of the candidate video are respectively input into a pre-trained first sorting model and a second sorting model to obtain a first sorting result and a second sorting result of the candidate video.

[0037] The positive and negative training samples of the second sorting model are determined according to a pre-trained third sorting model, and the sorting accuracy of the third sorting model is higher than the sorting accuracy of the first sorting model and the second sorting model.

[0038] In implementation, a pre-trained first sorting model and a second sorting model may be stored in the video server. Among them, the positive and negative training samples of the second sorting model are determined according to the pre-trained third sorting model; the sorting accuracy of the third sorting model is higher than the sorting accuracy of the first sorting model and the second sorting model. After the video server obtains the user features of the target user and the video features of each candidate video, for each candidate video in each candidate video, the video server can input the user features of the target user and the video features of the candidate video into the pre-trained first sorting model and the second sorting model respectively. Correspondingly, the first sorting model will output the first sorting result of the candidate video, and the second sorting model will output the second sorting result of the candidate video. Optionally, the training method of the second sorting model will be described in detail later.

[0039] Step 103, determining a fusion ranking result of each candidate video according to the first ranking result and the second ranking result of each candidate video, and selecting a first preset number of candidate videos from each candidate video in descending order of the fusion ranking result as target videos for the target user.

[0040] In implementation, the weight of the first sorting model and the weight of the second sorting model may be pre-stored in the video server. The weight of the first sorting model and the weight of the second sorting model may be set by a technician based on experience. After the video server obtains the first sorting result and the second sorting result of the candidate video, it may calculate the weighted sum of the first sorting result and the second sorting result of the candidate video, and use the weighted sum as the fusion sorting result of the candidate video. For example, the first sorting result of the candidate video is 6, the second sorting result is 8, the weight of the first sorting model is 0.4, and the weight of the second sorting model is 0.6, then the fusion sorting result of the candidate video is 6*0.4+8*0.6=7.2.

[0041] The video server may store a first preset number in advance, and the first preset number may be set by a technician based on experience. After obtaining the fusion ranking results of each candidate video, the video server may select the first preset number of candidate videos from each candidate video in descending order of the fusion ranking results as the target videos of the target user.

[0042] Step 104: for each target video in each target video, the user features of the target user and the video features of the target video are input into a third sorting model to obtain a third sorting result of the target video.

[0043] In implementation, a pre-trained third sorting model may be stored in the video server. After the video server obtains the target video of the target user, for each target video in the target videos, the video server may input the user features of the target user and the video features of the target video into the pre-trained third sorting model respectively. Correspondingly, the third sorting model will output the third sorting result of the target video.

[0044] Step 105: recommend videos to target users based on the third ranking results of the target videos.

[0045] In implementation, after the video server obtains the third ranking result of each target video, it can recommend the video to the target user according to the preset video recommendation algorithm. In this way, since the training process of the rough ranking model and the fine ranking model are related to each other, and the positive and negative training samples used by the second ranking model are determined by the third ranking model, it is ensured that the rough ranking stage will not filter out the videos with higher scores in the fine ranking stage among the candidate videos, thereby improving the accuracy of the target video recommended to the user.

[0046] Figure 2 is a flowchart of a training method for a second sorting model according to an exemplary embodiment. Figure 2 As shown, the method may include the following steps.

[0047] Step 201: Acquire multiple training samples, wherein the training samples include user features of sample users and video features of sample videos.

[0048] In implementation, the video server may pre-store the second ranking model to be trained and a plurality of training samples. The training samples include user features of sample users and video features of sample videos. When the video server needs to train the second ranking model, the video server may obtain a plurality of pre-stored training samples. Optionally, the user features of the sample users may also include the estimated click-through rate, like rate, and attention rate of the sample users obtained by the first ranking model; the video features of the sample videos may also include the estimated click-through rate, like rate, and attention rate of the sample videos obtained by the first ranking model.

[0049] Step 202: for each training sample, the user features of the sample user and the video features of the sample video contained in the training sample are input into a pre-trained third sorting model to obtain a fourth sorting result of the sample videos in the training sample.

[0050] In implementation, the video server may store a pre-trained third ranking model. After the video server obtains multiple training samples, for each training sample, the video server may input the user features of the sample user and the video features of the sample video contained in the training sample into the pre-trained third ranking model. Correspondingly, the third ranking model will output the fourth ranking result of the sample video in the training sample.

[0051] Step 203: determine positive training samples and negative training samples according to the fourth ranking result of the sample videos in each training sample.

[0052] In implementation, after obtaining the fourth ranking result of the sample videos in each training sample, the video server may determine the positive training sample and the negative training sample according to the fourth ranking result of the sample videos in each training sample.

[0053] Optionally, the video server may determine the positive training samples and the negative training samples in various ways according to the fourth ranking result of the sample videos in each training sample. The embodiment of the present disclosure provides two feasible implementations, and the specific processing process is as follows.

[0054] In implementation mode 1, the video server may select a second preset number of first training samples from each training sample in descending order of the fourth sorting result as positive training samples, and select the second training samples in each training sample except the first training samples as negative training samples.

[0055] In implementation, a second preset number may be pre-stored in the video server. The second preset number may be set by a technician based on experience. After the video server obtains the fourth sorting result of the sample videos in each training sample, for each sample user, the video server may select a second preset number of first training samples from each training sample corresponding to the sample user in descending order of the fourth sorting result as positive training samples, and use the second training samples in each training sample except the first training sample as negative training samples. In this way, the video server determines the fourth sorting result of each training sample through the third sorting model. Then, the video server selects the positive and negative training samples used by the second sorting model according to the fourth sorting result of each training sample, thereby ensuring that the rough sorting stage will not filter out the videos with higher scores in the fine sorting stage among the candidate videos.

[0056] In the second implementation mode, the video server determines the processing process of positive training samples and negative training samples according to the fourth ranking result of each training sample as follows.

[0057] Step 1: Determine the recommended sorting result of the sample videos in each training sample according to a preset recommended sorting algorithm and the fourth sorting result of the sample videos in each training sample.

[0058] In implementation, the video server may pre-store a recommendation sorting algorithm corresponding to the strategy stage. The recommendation sorting algorithm may be set by a technician as needed. After the video server obtains the fourth sorting result of the sample videos in each training sample, for each sample user, the video server may determine the recommendation sorting result of the sample videos in each training sample corresponding to the sample user according to the preset recommendation sorting algorithm and the fourth sorting result of the sample videos in each training sample corresponding to the sample user.

[0059] Step 2: Select a third preset number of third training samples from each training sample in descending order of the fourth sorting results, select a fourth preset number of fourth training samples from each training sample in descending order of the recommended sorting results, and determine the union of the third training samples and the fourth training samples as the positive training samples.

[0060] In implementation, a third preset number and a fourth preset number may be pre-stored in the video server. The third preset number and the fourth preset number may be set by a technician based on experience. For each sample user, after the video server obtains the fourth sorting result and the recommended sorting result of the sample videos in each training sample corresponding to the sample user, it may further select a third preset number (e.g., 10) of third training samples from each training sample corresponding to the sample user in the order of the fourth sorting result from high to low. Similarly, the video server may select a fourth preset number (e.g., 10) of fourth training samples from each training sample corresponding to the sample user in the order of the recommended sorting result from high to low. Then, the video server may determine the union of the third training sample and the fourth training sample corresponding to the sample user as the positive training sample corresponding to the sample user.

[0061] Step three, in the order of the fourth sorting results from low to high, select the fifth preset number of fifth training samples from each training sample, in the order of the recommended sorting results from low to high, select the sixth preset number of sixth training samples from each training sample, and determine the union of the fifth training sample and the sixth training sample as the candidate negative training sample.

[0062] In implementation, the video server may pre-store a fifth preset number and a sixth preset number. The fifth preset number and the sixth preset number may be set by a technician based on experience. For each sample user, after the video server obtains the fourth sorting result and the recommended sorting result of the sample videos in each training sample corresponding to the sample user, it may further select the fifth preset number of fifth training samples from each training sample corresponding to the sample user in the order from low to high of the fourth sorting result. Similarly, the video server may select the sixth preset number of sixth training samples from each training sample corresponding to the sample user in the order from low to high of the recommended sorting result. Then, the video server may determine the union of the fifth training sample corresponding to the sample user and the sixth training sample corresponding to the sample user as the candidate negative training sample corresponding to the sample user.

[0063] Step 4: Among the candidate negative training samples, a preset proportion of second candidate negative training samples other than the first candidate negative training samples selected as positive training samples are used as negative training samples.

[0064] In implementation, a preset ratio may be pre-stored in the video server. The preset ratio may be set by a technician based on experience. For each sample user, after the video server obtains the candidate negative training sample corresponding to the sample user, it may select a preset ratio (e.g., 5%) of second candidate negative training samples other than the first candidate negative training sample that has been selected as the positive training sample as negative training samples in the candidate negative training samples. In this way, the video server determines the fourth sorting result of each training sample through the third sorting model. Then, the video server determines the recommended sorting result of the sample video in each training sample based on the recommended sorting algorithm and the fourth sorting result of the sample video in each training sample. Afterwards, the video server selects the positive and negative training samples used by the second sorting model based on the fourth sorting result of each training sample and the recommended sorting result, thereby ensuring that the rough sorting stage will not filter out the videos with higher scores in the candidate videos in the fine sorting stage and the strategy stage.

[0065] Step 204: train the second sorting model to be trained according to the positive training samples and the negative training samples to obtain a trained second sorting model.

[0066] In implementation, after the video server obtains the positive training samples and the negative training samples, it can train the second sorting model to be trained according to the positive training samples and the negative training samples to obtain the trained second sorting model. In this way, the video server determines the fourth sorting result of each training sample through the third sorting model. Then, the video server selects the positive and negative training samples used by the second sorting model according to the fourth sorting result of each training sample, thereby ensuring that the rough sorting stage will not filter out the videos with higher scores in the fine sorting stage among the candidate videos.

[0067] The disclosed embodiment provides a method for recommending videos. A video server obtains user features of a target user and video features of each candidate video. Then, for each candidate video in each candidate video for video recommendation, the video server inputs the user features of the target user for video recommendation and the video features of the candidate video into a pre-trained first sorting model and a second sorting model, respectively, to obtain a first sorting result and a second sorting result of the candidate video. Among them, the positive and negative training samples of the second sorting model for video recommendation are determined according to a pre-trained third sorting model, and the sorting accuracy of the third sorting model for video recommendation is higher than the sorting accuracy of the first sorting model for video recommendation and the second sorting model for video recommendation. Afterwards, the video server determines the fusion sorting results of each candidate video for video recommendation according to the first sorting results and the second sorting results of each candidate video for video recommendation, and selects a first preset number of candidate videos from each candidate video for video recommendation in the order of the fusion sorting results of video recommendation from high to low as the target videos for the target user for video recommendation. Finally, for each target video in each target video, the video server inputs the user features of the target user of the video recommendation and the video features of the target video into the third ranking model of the video recommendation, obtains the third ranking result of the target video, and recommends the video to the target user of the video recommendation according to the third ranking result of each target video of the video recommendation. In this way, since the training processes of the second ranking model and the third ranking model are interrelated, and the positive and negative training samples used by the second ranking model are determined by the third ranking model, it is ensured that the rough ranking stage will not filter out the videos with higher scores in the fine ranking stage among the candidate videos, thereby improving the accuracy of the target videos recommended for the user.

[0068] Figure 3 FIG. 1 is a block diagram of a video recommendation device according to an exemplary embodiment. Figure 3 , the device comprises:

[0069] A first acquisition module 310 is configured to acquire user features of a target user and video features of each candidate video;

[0070] The first input module 320 is configured to execute, for each candidate video in each candidate video, inputting the user features of the target user and the video features of the candidate video into the pre-trained first sorting model and the second sorting model respectively, to obtain the first sorting result and the second sorting result of the candidate video, wherein the positive and negative training samples of the second sorting model are determined according to the pre-trained third sorting model, and the sorting accuracy of the third sorting model is higher than the sorting accuracy of the first sorting model and the second sorting model;

[0071] The selection module 330 is configured to determine the fusion ranking result of each candidate video according to the first ranking result and the second ranking result of each candidate video, and select a first preset number of candidate videos from each candidate video in descending order of the fusion ranking result as the target videos of the target user;

[0072] The second input module 340 is configured to input the user features of the target user and the video features of the target video into the third sorting model for each target video in the target videos, so as to obtain a third sorting result of the target video;

[0073] The recommendation module 350 is configured to recommend videos to target users according to the third ranking results of the target videos.

[0074] As an optional implementation, Figure 4 As shown, the device also includes:

[0075] The second acquisition module 360 ​​is configured to acquire a plurality of training samples, where the training samples include user features of sample users and video features of sample videos;

[0076] The third input module 370 is configured to input, for each training sample, the user features of the sample user and the video features of the sample video contained in the training sample into a pre-trained third sorting model to obtain a fourth sorting result of the sample videos in the training sample;

[0077] A determination module 380 is configured to determine positive training samples and negative training samples according to a fourth ranking result of the sample videos in each training sample;

[0078] The training module 380 is configured to train the second sorting model to be trained according to the positive training samples and the negative training samples to obtain the trained second sorting model.

[0079] As an optional implementation, the determination module 380 is specifically configured to:

[0080] According to the fourth sorting result from high to low order, in each training sample, a second preset number of first training samples are selected as positive training samples, and the second training samples except the first training samples in each training sample are selected as negative training samples.

[0081] As an optional implementation, the determination module 380 is specifically configured to:

[0082] Determine a recommended sorting result of the sample videos in each training sample according to a preset recommended sorting algorithm and a fourth sorting result of the sample videos in each training sample;

[0083] According to the fourth sorting result from high to low order, a third preset number of third training samples are selected from each training sample, according to the recommended sorting result from high to low order, a fourth preset number of fourth training samples are selected from each training sample, and a union of the third training sample and the fourth training sample is determined as a positive training sample;

[0084] According to the fourth sorting result from low to high order, a fifth preset number of fifth training samples are selected from each training sample, according to the recommended sorting result from low to high order, a sixth preset number of sixth training samples are selected from each training sample, and a union of the fifth training sample and the sixth training sample is determined as a candidate negative training sample;

[0085] Among the candidate negative training samples, a preset proportion of second candidate negative training samples other than the first candidate negative training samples selected as positive training samples are used as negative training samples.

[0086] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0087] The disclosed embodiment provides a video recommendation device. A video server obtains user features of a target user and video features of each candidate video. Then, for each candidate video in each candidate video of the video recommendation, the video server inputs the user features of the target user of the video recommendation and the video features of the candidate video into a pre-trained first sorting model and a second sorting model, respectively, to obtain a first sorting result and a second sorting result of the candidate video. Among them, the positive and negative training samples of the second sorting model of the video recommendation are determined according to a pre-trained third sorting model, and the sorting accuracy of the third sorting model of the video recommendation is higher than the sorting accuracy of the first sorting model of the video recommendation and the second sorting model of the video recommendation. Afterwards, the video server determines the fusion sorting results of each candidate video of the video recommendation according to the first sorting results and the second sorting results of each candidate video of the video recommendation, and selects a first preset number of candidate videos from each candidate video of the video recommendation in the order of the fusion sorting results of the video recommendation from high to low as the target videos of the target user of the video recommendation. Finally, for each target video in each target video, the video server inputs the user features of the target user of the video recommendation and the video features of the target video into the third ranking model of the video recommendation, obtains the third ranking result of the target video, and recommends the video to the target user of the video recommendation according to the third ranking result of each target video of the video recommendation. In this way, since the training processes of the second ranking model and the third ranking model are interrelated, and the positive and negative training samples used by the second ranking model are determined by the third ranking model, it is ensured that the rough ranking stage will not filter out the videos with higher scores in the fine ranking stage among the candidate videos, thereby improving the accuracy of the target videos recommended for the user.

[0088] Figure 5 The block diagram of a video server according to an exemplary embodiment is shown. The video server comprises: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the above-mentioned video recommendation method.

[0089] In an exemplary embodiment, a storage medium including instructions is also provided, such as a memory including instructions, and the instructions can be executed by a processor of a video server to complete the above method. Optionally, the storage medium can be a non-transitory computer-readable storage medium, for example, the non-transitory computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0090] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0091] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A video recommendation method, characterized in that: The method comprises: Obtaining user features of the target user and video features of each candidate video; For each candidate video among the candidate videos, the user features of the target user and the video features of the candidate video are respectively input into a pre-trained first sorting model and a second sorting model to obtain a first sorting result and a second sorting result of the candidate video, wherein the positive and negative training samples of the second sorting model are determined according to a pre-trained third sorting model, and the sorting accuracy of the third sorting model is higher than the sorting accuracy of the first sorting model and the second sorting model; the user features include user ID, click-through rate, like rate, attention rate, and the ID, author ID, click-through rate, like rate and attention rate of the browsed video; the video features include the ID, author ID, click-through rate, like rate and attention rate of the video, and the user ID, click-through rate, like rate and attention rate of the user browsing the video; Determine a fusion ranking result of the candidate videos according to the first ranking result and the second ranking result of the candidate videos, and select a first preset number of candidate videos from the candidate videos in descending order of the fusion ranking results as target videos of the target user; For each target video in each target video, input the user feature of the target user and the video feature of the target video into the third sorting model to obtain a third sorting result of the target video; Recommending videos to the target user according to the third ranking results of the target videos.

2. The method according to claim 1, characterized in that: Before obtaining the user features of the target user and the video features of each candidate video, the method further includes: Acquire a plurality of training samples, wherein the training samples include user features of sample users and video features of sample videos; For each training sample, inputting user features of a sample user and video features of a sample video contained in the training sample into a pre-trained third ranking model to obtain a fourth ranking result of the sample videos in the training sample; Determine positive training samples and negative training samples according to the fourth sorting result of the sample videos in each training sample; The second sorting model to be trained is trained according to the positive training samples and the negative training samples to obtain a trained second sorting model.

3. The method according to claim 2, characterized in that The determining of positive training samples and negative training samples according to the fourth sorting result of the sample videos in each training sample includes: According to the fourth sorting result from high to low order, in each training sample, a second preset number of first training samples are selected as positive training samples, and second training samples except the first training samples in each training sample are selected as negative training samples.

4. The method according to claim 2, characterized in that: The determining of positive training samples and negative training samples according to the fourth sorting result of the sample videos in each training sample includes: Determine a recommended sorting result of the sample videos in each training sample according to a preset recommended sorting algorithm and a fourth sorting result of the sample videos in each training sample; According to the fourth sorting result from high to low order, a third preset number of third training samples are selected from each training sample, according to the recommended sorting result from high to low order, a fourth preset number of fourth training samples are selected from each training sample, and a union of the third training sample and the fourth training sample is determined as a positive training sample; According to the fourth sorting result from low to high order, a fifth preset number of fifth training samples are selected from the training samples, according to the recommended sorting result from low to high order, a sixth preset number of sixth training samples are selected from the training samples, and a union of the fifth training sample and the sixth training sample is determined as a candidate negative training sample; Among the candidate negative training samples, a preset proportion of second candidate negative training samples other than the first candidate negative training samples selected as the positive training samples are used as negative training samples.

5. A video recommendation device, characterized in that: The device comprises: A first acquisition module is configured to acquire user features of a target user and video features of each candidate video; A first input module is configured to execute, for each candidate video in the candidate videos, inputting the user features of the target user and the video features of the candidate video into a pre-trained first sorting model and a second sorting model respectively, to obtain a first sorting result and a second sorting result of the candidate video, wherein the positive and negative training samples of the second sorting model are determined according to a pre-trained third sorting model, and the sorting accuracy of the third sorting model is higher than the sorting accuracy of the first sorting model and the second sorting model; the user features include user identification, click-through rate, like rate, attention rate, and identification, author identification, click-through rate, like rate and attention rate of the browsed video; the video features include video identification, author identification, click-through rate, like rate and attention rate, and user identification, click-through rate, like rate and attention rate of the user browsing the video; A selection module is configured to determine a fusion ranking result of the candidate videos according to the first ranking result and the second ranking result of the candidate videos, and select a first preset number of candidate videos from the candidate videos in descending order of the fusion ranking results as target videos of the target user; A second input module is configured to input the user features of the target user and the video features of the target video into the third sorting model for each target video in the target videos, so as to obtain a third sorting result of the target video; The recommendation module is configured to recommend videos to the target user according to the third ranking results of the target videos.

6. The device according to claim 5, characterized in that The device also includes: A second acquisition module is configured to acquire a plurality of training samples, wherein the training samples include user features of sample users and video features of sample videos; A third input module is configured to input, for each training sample, user features of a sample user and video features of a sample video contained in the training sample into a pre-trained third sorting model to obtain a fourth sorting result of the sample videos in the training sample; A determination module is configured to determine positive training samples and negative training samples according to a fourth sorting result of the sample videos in each training sample; The training module is configured to train the second sorting model to be trained according to the positive training samples and the negative training samples to obtain the trained second sorting model.

7. The device according to claim 6, characterized in that The determination module is specifically used for: According to the fourth sorting result from high to low order, in each training sample, a second preset number of first training samples are selected as positive training samples, and second training samples except the first training samples in each training sample are selected as negative training samples.

8. The device according to claim 6, characterized in that The determination module is specifically used for: Determine a recommended sorting result of the sample videos in each training sample according to a preset recommended sorting algorithm and a fourth sorting result of the sample videos in each training sample; According to the fourth sorting result from high to low order, a third preset number of third training samples are selected from each training sample, according to the recommended sorting result from high to low order, a fourth preset number of fourth training samples are selected from each training sample, and a union of the third training sample and the fourth training sample is determined as a positive training sample; According to the fourth sorting result from low to high order, a fifth preset number of fifth training samples are selected from the training samples, according to the recommended sorting result from low to high order, a sixth preset number of sixth training samples are selected from the training samples, and a union of the fifth training sample and the sixth training sample is determined as a candidate negative training sample; Among the candidate negative training samples, a preset proportion of second candidate negative training samples other than the first candidate negative training samples selected as the positive training samples are used as negative training samples.

9. A video server, characterized in that: The video server comprises: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the video recommendation method according to any one of claims 1 to 4.

10. A storage medium, when instructions in the storage medium are executed by a processor of a video server, the video server is enabled to execute the video recommendation method according to any one of claims 1 to 4.

Citation Information

Patent Citations

  • Video recommendation method and device, electronic equipment and storage medium

    CN110019943A