WO2020135054A1 - Procédé, dispositif et appareil de recommandation de vidéos et support de stockage - Google Patents

Procédé, dispositif et appareil de recommandation de vidéos et support de stockage Download PDF

Info

Publication number
WO2020135054A1
WO2020135054A1 PCT/CN2019/124582 CN2019124582W WO2020135054A1 WO 2020135054 A1 WO2020135054 A1 WO 2020135054A1 CN 2019124582 W CN2019124582 W CN 2019124582W WO 2020135054 A1 WO2020135054 A1 WO 2020135054A1
Authority
WO
WIPO (PCT)
Prior art keywords
video
recommended
historical
classification
recommendation model
Prior art date
Application number
PCT/CN2019/124582
Other languages
English (en)
Chinese (zh)
Inventor
刘运
刘文奇
Original Assignee
广州市百果园信息技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 广州市百果园信息技术有限公司 filed Critical 广州市百果园信息技术有限公司
Publication of WO2020135054A1 publication Critical patent/WO2020135054A1/fr

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
    • H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
    • H04N21/466—Learning process for intelligent management, e.g. learning user preferences for recommending movies
    • H04N21/4668—Learning process for intelligent management, e.g. learning user preferences for recommending movies for recommending content, e.g. movies
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/20—Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
    • H04N21/25—Management operations performed by the server for facilitating the content distribution or administrating data related to end-users or client devices, e.g. end-user or client device authentication, learning user preferences for recommending movies
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/20—Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
    • H04N21/25—Management operations performed by the server for facilitating the content distribution or administrating data related to end-users or client devices, e.g. end-user or client device authentication, learning user preferences for recommending movies
    • H04N21/251—Learning process for intelligent management, e.g. learning user preferences for recommending movies
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
    • H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
    • H04N21/466—Learning process for intelligent management, e.g. learning user preferences for recommending movies
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
    • H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
    • H04N21/466—Learning process for intelligent management, e.g. learning user preferences for recommending movies
    • H04N21/4662—Learning process for intelligent management, e.g. learning user preferences for recommending movies characterized by learning algorithms
    • H04N21/4665—Learning process for intelligent management, e.g. learning user preferences for recommending movies characterized by learning algorithms involving classification methods, e.g. Decision trees
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
    • H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
    • H04N21/466—Learning process for intelligent management, e.g. learning user preferences for recommending movies
    • H04N21/4667—Processing of monitored end-user data, e.g. trend analysis based on the log file of viewer selections

Definitions

  • the embodiments of the present application relate to computer vision technology, for example, to a video recommendation method, device, device, and storage medium.
  • Video recommendation is an important research direction in the field of computer vision.
  • deep learning is also widely used in video recommendation, that is, a video recommendation model based on deep learning training is used for video recommendation.
  • the video recommendation model needs to be adjusted according to actual business needs.
  • the video recommendation model generated based on deep learning training has a long training period, which makes it difficult to adapt to the adjustment of actual services.
  • the adjustment periods of the above actual services are relatively short, making deep learning based
  • the prediction accuracy of the video recommendation model generated by the training is not high, and it is impossible to obtain a suitable video for recommendation.
  • the adjustment of the actual business can refer to the adjustment of the operation strategy or the change of the hotspot video.
  • the so-called operation strategy adjustment can be understood as changing the recommended animation video to the recommended game video, and the change of the hotspot video can be understood as the change of the hotspot video from category C to category D.
  • the recommended category C video becomes the recommended category D video.
  • Embodiments of the present application provide a video recommendation method, device, equipment, and storage medium to improve the prediction accuracy of a video recommendation model.
  • An embodiment of the present application provides a video recommendation method.
  • the method includes:
  • the recommended video under each classification label is obtained.
  • An embodiment of the present application also provides a video recommendation device, which includes:
  • the current video collection acquisition module is set to obtain the current video collection
  • the classification label and classification score acquisition module is set to input the current video collection and the historical video collection into the historical latest video recommendation model to obtain the classification label and classification score of the video in the current video collection;
  • the recommended video determination module is configured to obtain the recommended video under each classification label according to the classification score of the video under each classification label in the current video set.
  • An embodiment of the present application also provides a device, which includes:
  • One or more processors are One or more processors;
  • Memory set to store one or more programs
  • the one or more processors When the one or more programs are executed by the one or more processors, the one or more processors implement the video recommendation method as provided in any embodiment of the present application.
  • An embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, a video recommendation method as provided in any embodiment of the present application is implemented.
  • FIG. 1 is a flowchart of a video recommendation method provided by an embodiment of the present application
  • FIG. 2 is a schematic structural diagram of a video recommendation device provided by an embodiment of the present application.
  • FIG. 3 is a schematic structural diagram of a device provided by an embodiment of the present application.
  • FIG. 1 is a flowchart of a video recommendation method provided by an embodiment of the present application. This embodiment can be applied to improve the prediction accuracy of a video recommendation model.
  • the method can be performed by a video recommendation device, which can use software and And/or hardware, the device may be configured in a device, such as a computer or a mobile terminal. As shown in Figure 1, the method includes the following steps:
  • Step 110 Obtain the current video collection.
  • Step 120 Input the current video collection and the historical video collection into the historical latest video recommendation model, and obtain the classification label and classification score of the video in the current video collection.
  • the current video collection may be a collection composed of videos without classification tags.
  • the current video collection may be a collection composed of videos uploaded on the same day without a classification label.
  • the historical video set may be a set composed of videos with classification tags, that is, each video in the historical video set has a corresponding classification tag.
  • the video classification label may refer to the content category corresponding to the video content, that is, the video classification label may refer to the content category to which the video belongs.
  • the historical video collection is a collection composed of videos with classification tags
  • the historical video collection and the current video collection are input into the historical latest video recommendation model together, and the historical video collection can be used as a priori knowledge of the current video collection, thereby improving the history The prediction accuracy of the latest video recommendation model for the current video collection.
  • the historical latest video recommendation model can be used to determine the corresponding classification label for each video in the current video collection.
  • the network structure of the latest historical video recommendation model can be the network structure of the deep learning algorithm, such as the network structure of the traditional convolutional neural network, or the lightweight structure formed by improving the network structure of the traditional convolutional neural network.
  • the latest historical video recommendation model may be a video recommendation model that is updated according to a preset period. The update described here is to calculate the loss function of the model through forward propagation according to the updated input variables, and calculate the loss function on the network parameters of the model. Partial derivative, and then update the network parameters of the model through reverse gradient propagation until the new network parameters are determined.
  • the historical latest video recommendation model may be a video recommendation model updated according to a preset update period
  • the historical latest video recommendation model may refer to the video recommendation model corresponding to the current update period.
  • the source of the updated input variable described above may be determined based on the prediction result of the video recommendation model corresponding to the last update cycle.
  • the historical latest video recommendation model described in the embodiment of the present application is different from the video recommendation model generated based on deep learning training in the related art.
  • the video recommendation model that can be applied to the actual business in the related technology often needs to go through the following three stages: First stage, collect and label data; second stage, training model; third stage, input the data from the actual online business into the trained video recommendation model, and evaluate the prediction accuracy of the model according to the prediction results of the video recommendation model Then, according to the evaluation results, adjust the input variables and retrain the model until the evaluation results meet the actual business application requirements.
  • the data in the first stage refers to offline data.
  • the offline data is input to the model as input variables. If the actual business changes, the above three stages need to be re-executed to obtain a video recommendation model corresponding to the actual business.
  • the foundation of the historical latest video recommendation model described in the embodiments of the present application is a pre-training model, which is a multi-modal model that has learned a large number of related data sets.
  • the multi-modal model includes an audio recognition model and an action recognition model , Image classification model, image segmentation model and face recognition model.
  • the video recommendation model that can be applied to the actual business in the related technologies cannot achieve good results in recommending hot videos. This is due to the long period of obtaining the video recommendation model that can be applied to the actual business in the related technology, and the update period of the hotspot video is relatively short.
  • the historical latest video recommendation model provided by the embodiment of the present application can achieve better results in recommending hotspot videos. This is because the embodiment of the present application has a shorter period of obtaining the historical latest video recommendation model that can be applied to actual services. It can be matched with the update period of hotspot videos.
  • the historical latest video recommendation model is a video recommendation model updated according to the update cycle, the historical latest video recommendation model can adapt to the changes of the hotspot video in real time and make corresponding changes to recommend the hotspot video.
  • the original classification label of the video may be sorted according to the descending sorting result of the original classification score of the video, and the top N original classification labels are selected from the sorting results of the original classification labels as The category label for this video.
  • each original classification label corresponds to an original classification score.
  • the historical latest video recommendation model described in the embodiment of the present application is a multi-category video recommendation model, that is, the number of original classification labels of each video in the current video set obtained based on the historical latest video recommendation model is two or two More than.
  • each original classification label corresponds to an original classification score, that is, the number of original classification labels is the same as the original classification score
  • the difference is that the original classification score corresponding to each original classification label, of course, may also exist
  • the number of original classification tags and the number of original classification tags are determined by the historical latest video recommendation model.
  • the number of original classification tags that can be recognized by the latest historical video recommendation model is three, and the three original classification tags are music, basketball, and original
  • the number of original classification tags of video A in the current video collection That is three, namely music, basketball and original, the original classification scores corresponding to the three original classification tags of video A are 5, 4 and 1, respectively
  • the number of original classification tags of video B is also three, respectively Music, basketball and original, the original classification scores corresponding to the three original classification labels of video B are 1, 2 and 6, respectively.
  • For video A sort the original classification tags of video A according to the descending sorting results of the original classification score of video A, and select the first two original classification tags from the sorting results of the original classification tags of video A as the classification of video A Tag, that is, the original classification label of video A, music and basketball as the classification label of video A; for video B, the original classification label of video B is sorted according to the descending sorting result of the original classification score of video B, and The first two original classification tags are selected as the classification tags of video B in the sorting results of the original classification tags, that is, the original classification tags of video B, basketball and original, are used as the classification tags of video B.
  • Step 130 According to the classification score of the video under each classification label in the current video set, obtain the recommended video under each classification label.
  • the recommended video under each classification label needs to be determined. That is, for each classification label, multiple videos can be sorted according to the classification score corresponding to multiple videos in the current video set, and the recommended videos under the classification label can be determined according to the sorting result. In an embodiment, for each classification label, multiple videos can be sorted according to the descending sorting results of the classification scores corresponding to multiple videos in the current video set, and the top M videos can be selected from the sorting results of multiple videos As a recommended video under this category label. Each category label performs the same operation as described above until the recommended video under each category label is determined. In an embodiment, the value of M corresponding to different classification labels may be the same or different, and may be determined according to actual conditions, which is not limited herein.
  • obtaining the recommended video under each classification label according to the classification score of the video under each classification label in the current video collection may include: according to each in the current video collection The classification score of the video under the classification label to obtain the video to be recommended under each classification label.
  • the classification label of the video to be recommended is consistent with the content of the video to be recommended, the video to be recommended is regarded as the recommended video under each classification label.
  • the original classification score corresponding to the original classification label of each video in the current video set obtained based on the historical latest video recommendation model may not be correct, based on the original classification score of each video, it is determined
  • the classification label of each video may also be incorrect.
  • the classification score corresponding to the classification label of the video is not correct.
  • the recommended The video may also be incorrect.
  • the following method may be adopted: according to the classification scores of multiple videos under each classification label in the current video set, the videos to be recommended under each classification label are obtained, and the classification of the videos to be recommended is determined Whether the label is consistent with the content of the video to be recommended. If the classification label of the video to be recommended is consistent with the content of the video to be recommended, it can indicate that the recommended video under each classification label obtained based on the historical latest video recommendation model is correct.
  • the to-be-recommended video determined for each category label serves as the recommended video under each category label.
  • the above determination of whether the classification label of the video to be recommended is consistent with the content of the video to be recommended may be performed by a user who has permission to review the classification label of the video.
  • the current video collection and the historical video collection are input into the historical latest video recommendation model, and the classification labels and classification scores of the videos in the current video collection are obtained, according to each classification label in the current video collection The classification score of the next video to obtain the recommended video under each classification label.
  • the latest historical video recommendation model has a shorter training period than the video recommendation model generated based on deep learning training in the related technology, the historical video The recommendation model can be well adapted to the adjustment of actual business.
  • the historical video collection that can be used as a priori knowledge of the current video collection is also used as an input variable, which improves the prediction accuracy of the latest video recommendation model for the current video collection.
  • the method may further include: when the classification label of the video to be recommended is different from the content of the video to be recommended, after the preset time point is reached, the video to be recommended will be added
  • the historical video collection is used as the input variable of the historical latest video recommendation model to update the historical latest video recommendation model.
  • the classification label of the video to be recommended is not consistent with the content corresponding to the content of the video to be recommended, it may indicate that the recommended video under the classification label obtained based on the historical latest video recommendation model is incorrect, and at the same time .
  • the latest historical video recommendation model is the video recommendation model corresponding to the current update cycle, in other words, the video recommendation model is a video recommendation model that is updated according to the update cycle, therefore, the video recommendation model needs to be updated to get the next update
  • the historical latest video recommendation model corresponding to the period improves the prediction accuracy of the historical latest video recommendation model corresponding to the next update period, and the to-be-recommended videos whose classification labels of the to-be-recommended videos and the content of the to-be-recommended videos are inconsistent can be added to the historical video collection ,
  • the historical video set added to the to-be-recommended video is used as an input variable of the historical latest video recommendation model, and participates in the process of updating the historical latest video recommendation model.
  • the reason why the above can improve the prediction accuracy of the historical latest video recommendation model corresponding to the next update cycle is that: the category label of the above-mentioned to-be-recommended video and the content of the to-be-recommended video are inconsistent, and it is considered that the latest video recommendation model is not easy to be accurate Determine the video of the classification label, and input the historical video set added to the video to be recommended as the input variable into the historical latest video recommendation model, so that the historical latest video recommendation model can further learn the characteristics of the video to be recommended during the update process to improve
  • the recognition accuracy of the classification label of the video to be recommended by the historical latest video recommendation model corresponding to the next update cycle also increases the prediction accuracy of the historical latest video recommendation model corresponding to the next update cycle.
  • a video in which the classification label of the video to be recommended is inconsistent with the content corresponding to the content of the video to be recommended is called a hard sample, that is, a hard sample is a video of the following type: a video in the current video set is originally A video that does not belong to a category label but is labeled by the latest historical video recommendation model with a high category score as the category label.
  • determining whether the classification label marked on the video is correct can be performed by a user who has permission to review the classification label of the video.
  • the method may further include: obtaining a target video that meets the recommended conditions in the current video collection, and sending the target video to a terminal having a labeling permission for the classification label; receiving the label classification sent by the terminal The target video after the label; after reaching the preset time point, the historical video collection of the target video after adding the classification label is used as the input variable of the historical latest video recommendation model to update the historical latest video recommendation model.
  • the target video in the current video collection that meets the recommended conditions can also be obtained and the target video can be sent to
  • the corresponding labeling personnel will label the classification label and receive the target video after the labeling of the classification label.
  • the historical video collection of the target video with the labeling classification label will be added.
  • the historical video collection of the target video after adding the classification label is used as the input variable of the latest historical video recommendation model.
  • the reason for participating in the process of updating the historical latest video recommendation model is that the target video is sent to the label with the classification label
  • the corresponding tagging personnel tag the target videos in the current video set that meet the recommended conditions, and it is generally considered that the tagging results obtained by tagging the target videos in the above manner (that is, the target video’s Classification annotation) is relatively correct, therefore, the historical video collection of the target video after adding the classification label is used as the input variable of the latest historical video recommendation model, and participates in the process of updating the historical latest video recommendation model, making the latest historical video
  • the recommendation model can further strengthen the learning of the characteristics of the target video, and on the other hand, improve the prediction accuracy of the historical latest video recommendation model corresponding to the next update cycle.
  • the target video history video set with the classification label added is used as the input variable of the latest historical video recommendation model and participates in the process of updating the historical latest video recommendation model, then the distance needs to be deleted from the historical video collection
  • the target video after the classification label is marked in an update cycle with the longest update cycle time.
  • the recommended conditions may include a video playback rate greater than or equal to the playback rate threshold, a video like rate greater than or equal to the like rate threshold, a video comment rate greater than or equal to the review rate threshold, and a video forward rate greater than or equal to the forward rate threshold
  • the recommended condition is that the video play rate is greater than or equal to the play rate threshold
  • the target video that meets the recommended conditions refers to the video play rate greater than or equal to the play Rate threshold video
  • the recommended condition is that the video like rate is greater than or equal to the like rate threshold
  • the target video that meets the recommended conditions refers to the video whose video like rate is greater than or equal to the like rate threshold
  • the recommended condition is video If the review rate is greater than or equal to the review rate threshold, the target video that meets the recommended conditions refers to the video whose review rate is greater than or equal to the review rate threshold; if the recommended condition is that the video
  • the recommended conditions are two or more of the above, they must all be met.
  • the target video that meets the recommended conditions refers to the video play rate being greater than or equal to the play rate threshold and Videos with a video like rate greater than or equal to the like rate threshold; if the recommended conditions are that the video play rate is greater than or equal to the play rate threshold, the video like rate is greater than or equal to the like rate threshold, and the video forwarding rate is greater than the forward rate threshold, the recommended conditions are met
  • the target video refers to videos with a video playback rate greater than or equal to the playback rate threshold, a video like rate greater than or equal to the like rate threshold, and a video forward rate greater than or equal to the forward rate threshold.
  • the historical video collection of the target video after adding the classification label is used as the input variable of the historical latest video recommendation model to update the historical latest video recommendation
  • the model may include: assigning different weight values to the first video that is the same as the recommended video in the target video after the classification label and the second video that is not the same as the recommended video in the target video after the classification label.
  • the historical video collection of the first video after the weighting value and the second video after the weighting value are added is used as an input variable of the historical latest video recommendation model to update the historical latest video recommendation model.
  • the target video is a video that meets the recommendation condition in the current video collection, and the recommendation condition is a reflection of the user's real behavior
  • the target video is a video that needs to be labeled with a classification label. If the prediction accuracy of the historical latest video recommendation model is high, the target video should be included in the recommended video. In an embodiment, if the target video is included in the recommended video, the original video corresponding to the target video needs to have a relatively high score, so that it may be determined as the recommended video, and the prediction accuracy of the above-mentioned historical latest video recommendation model can only be achieved.
  • the recommended video does not contain the target video, it can indicate that the prediction accuracy of the historical latest video recommendation model is relatively low, and the original classification score of the target video cannot be accurately determined, which results in the historical latest video recommendation model being unable to correctly determine the target.
  • the reason for the original classification score of the video is that the latest video recommendation model cannot extract the effective features of the target video.
  • the target video that does not belong to the recommended video can be added to the historical video collection to participate in the update of the historical latest video recommendation model In the process, the historical latest video recommendation model can further strengthen the learning of the characteristics of the target video.
  • the historical latest video recommendation model can accurately determine the original classification score of the target video.
  • you can The target video of the recommended video and the target video that is not the recommended video are given different weight values. For the weight values corresponding to the two, a smaller weight value can be set for the target video that belongs to the recommended video and the target video that is not the recommended video Set a larger weight value.
  • the target video belonging to the recommended video means the video that appears in both the target video and the recommended video. From the perspective of the target video, the target video belonging to the recommended video can be understood as the first video in the target video that is the same as the recommended video, not The target video that belongs to the recommended video means a video that does not appear in the recommended video. From the perspective of the target video, the target video that does not belong to the recommended video can be understood as the second video in the target video that is different from the recommended video.
  • the target video that belongs to the recommended video and the target video that does not belong to the recommended video are given different weight values, that is, the first video in the target video that is the same as the recommended video and the second video in the target video that is not the same as the recommended video are given different weights Weights.
  • the weight value here indicates that in the process of updating the latest video recommendation model in history, when calculating the loss function of the video recommendation model, the loss function corresponding to the first video and the loss function corresponding to the second video are included in the loss function of the video recommendation model. Weight.
  • the first video in the target video that is the same as the recommended video may be called a positive sample.
  • the classification label corresponding to the positive sample is obtained by a person who has the authority to label the classification label.
  • the second video that is different from the recommended video in the above target video can be called a hard positive sample, that is, a hard positive sample refers to a video of the following type: a video in the current video collection should actually be a recommended video under a category label, However, because the latest historical video recommendation model processes the video, the video is given a lower original classification score, so that the video cannot be used as a recommended video under the classification label because the original classification score is low.
  • the second video in the target video after the classification label is different from the recommended video is used as the additional recommended video.
  • the second video since the second video does not belong to the recommended video, and the second video is a video that meets the recommendation condition in the current video collection, the second video that is different from the recommended video in the target video may be used as an additional recommendation video.
  • the method before inputting the current video collection and the historical video collection into the historical latest video recommendation model, the method may further include: acquiring the original historical video collection. According to the classification labels of the videos in the original historical video collection, cluster processing is performed to obtain the historical video collection.
  • the classification tags of the videos in the original historical video set may be clustered, and according to the clustering result, from each Under the category label, select the appropriate number of videos.
  • the reason for performing the above operation is that if the classification labels of the videos in the original historical video collection are not clustered, that is, the original historical video collection is used as the historical video collection, the number of videos under different classification labels in the historical video collection may not be Balanced, that is, the number of videos under one or more classification tags is relatively large, and the number of videos under other classification tags is relatively small.
  • the historical video collection will participate in the process of updating the latest video recommendation model, If the number of videos under different classification labels in the historical video collection is not balanced, it will result in the updated historical latest video recommendation model cannot accurately give the original classification score of the video with a small number of videos under the classification label. In short, the prediction accuracy of the updated historical latest video recommendation model will be low.
  • the number of videos under different classification tags in the historical video collection is more balanced, and the prediction accuracy of the updated historical latest video recommendation model is improved.
  • the number of original classification tags and original classification tags in the historical latest video recommendation model corresponding to the next update cycle is determined by the historical video set corresponding to the current update cycle. In one embodiment, the number of original classification tags and original classification tags in the historical latest video recommendation model corresponding to the next update cycle is determined by the first video in the target video corresponding to the recommended video and the recommended video in the target video corresponding to the current update cycle. The second video that is different from the video, and the first video that is the same as the recommended video in the target video corresponding to T update cycles before the current update cycle and the second video that is not the same as the recommended video in the target video. In terms of update cycle, the original classification labels and the number of original classification labels of the latest historical video recommendation model are also constantly updated.
  • the current video set corresponding to the current update cycle may be recommended
  • the classification label of the video is inconsistent with the content of the video to be recommended (i.e. difficult to sample), the first video (i.e.
  • the current update period corresponds The second video of the target video in the current video collection that is not the same as the recommended video (that is, difficult to positive samples), the first video that is the same as the recommended video in the target video corresponding to T update cycles before the current update cycle (that is, positive samples ) And in the target video corresponding to T update cycles before the current update cycle, different weight values are set for the second video (that is, difficult to sample) that is different from the recommended video.
  • the weight value here also represents the weight of the loss function corresponding to each of the above parts in the loss function of the video recommendation model when calculating the loss function of the video recommendation model during the update of the latest video recommendation model in history.
  • the video in the original historical video set is clustered, and only the first video in the target video that is the same as the recommended video in the current video set corresponding to the current update cycle and the target video may be different from the recommended video And the second video of the target video corresponding to the recommended video and the second video of the target video that is not the same as the recommended video are clustered.
  • the historical video set corresponding to the current update cycle will include the classification label determination of each video in the current video set corresponding to the previous update cycle.
  • Multi-category videos that is, videos whose classification labels of the videos to be recommended are inconsistent with the content of the videos to be recommended (that is, difficult samples), the first video in the target video that is the same as the recommended video (that is, positive samples), and the target video that is recommended
  • the first video that is, positive samples
  • the second video that is, difficult to positive samples
  • it also includes the first video (that is, positive samples) that is the same as the recommended video in the target video corresponding to T update cycles before the last update cycle and the target video that is the same as the recommended video.
  • the second video that is different from the recommended video that is, the difficult sample).
  • the target video is the same as the recommended video.
  • a video (that is, a positive sample) and a second video (that is, a difficult positive sample) that is different from the recommended video in the target video, and the target video that is the same as the recommended video in the target video corresponding to T update cycles before the last update cycle A video (that is, positive sample) and a second video (that is, difficult to positive sample) in the target video that is different from the recommended video will have a video with the classification label as Songkran.
  • the original classification label of the video is the original classification score of the video for the Songkran Festival It will be relatively high, so that the original classification label for the Songkran Festival is called the classification label of the video, and further, the video to be recommended under the classification label is obtained. If the classification label of the video to be recommended is consistent with the content of the video to be recommended, then The video to be recommended is regarded as the recommended video under the classification label, that is, the video whose classification label is Songkran Festival becomes the recommended video.
  • the video about the Songkran Festival is considered to be a recent hot video
  • FIG. 2 is a schematic structural diagram of a video recommendation device according to an embodiment of the present application. This embodiment can be applied to improve the prediction accuracy of a video recommendation model.
  • the device can be implemented in software and/or hardware.
  • the device can It is configured in the device, for example, in a computer or mobile terminal.
  • the device includes: a current video collection obtaining module 210, which is configured to obtain a current video collection.
  • the classification label and classification score acquisition module 220 is set to input the current video collection and the historical video collection into the historical latest video recommendation model to obtain the classification label and classification score of the video in the current video collection.
  • the recommended video determination module 230 is set to obtain the recommended video under each classification label according to the classification score of the video under each classification label in the current video set.
  • the current video collection and the historical video collection are input into the historical latest video recommendation model, and the classification labels and classification scores of the videos in the current video collection are obtained, according to each classification label in the current video collection
  • the classification score of the next video is used to obtain the recommended video under each classification label, and the video recommendation is performed by using the historical latest video recommendation model obtained on the basis of the pre-trained model.
  • the historical latest video recommendation model is based on The training period of the video recommendation model generated by deep learning training is short, so the historical video recommendation model can be well adapted to the adjustment of actual business, and at the same time, the historical video collection that can be used as a priori knowledge of the current video collection is also used as input Variables to improve the prediction accuracy of the latest video recommendation model for the current video collection.
  • FIG. 3 is a schematic structural diagram of a device provided by an embodiment of the present application.
  • FIG. 3 shows a block diagram of an exemplary device 312 suitable for implementing embodiments of the present application.
  • the device 312 is represented in the form of a general-purpose computing device.
  • the components of device 312 may include, but are not limited to, one or more processors 316, system memory 328, and bus 318 connected to different system components (including system memory 328 and processor 316).
  • System memory 328 may include computer system readable media in the form of volatile memory, such as random access memory (Random Access Memory, RAM) 330 and/or cache memory 332.
  • RAM random access memory
  • the storage system 334 may be configured to read and write non-removable, non-volatile magnetic media (not shown in FIG. 3, commonly referred to as a "hard disk drive").
  • the program/utility tool 340 having a set of (at least one) program modules 342 may be stored in the memory 328, for example.
  • the device 312 may also communicate with one or more external devices 314 (eg, keyboard, pointing device, display 324, etc.). This communication may be performed through an input/output (I/O) interface 322.
  • I/O input/output
  • the device 312 can also communicate with one or more networks (such as a local area network (Local Area Network, LAN), a wide area network (Wide Area Network, WAN), and/or a public network, such as the Internet) through the network adapter 320.
  • the processor 316 runs a program stored in the system memory 328 to execute various functional applications and data processing, for example, to implement a video recommendation method provided by an embodiment of the present application, the method includes: acquiring a current video collection. Input the current video collection and the historical video collection into the historical latest video recommendation model to obtain the classification label and classification score of the video in the current video collection. According to the classification score of the video under each classification label in the current video set, the recommended video under the classification label is obtained.
  • networks such as a local area network (Local Area Network, LAN), a wide area network (Wide Area Network, WAN), and/or a public network, such as the Internet
  • the processor 316 runs a program stored in the system memory 328 to execute various functional applications and
  • An embodiment of the present application also provides a computer-readable storage medium that stores a computer program, and when the program is executed by a processor, a video recommendation method as provided in the embodiment of the present application is implemented.
  • the method includes: obtaining a current video collection . Input the current video collection and the historical video collection into the historical latest video recommendation model to obtain the classification label and classification score of the video in the current video collection. According to the classification score of the video under each classification label in the current video set, the recommended video under the classification label is obtained.

Landscapes

  • Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computing Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Two-Way Televisions, Distribution Of Moving Picture Or The Like (AREA)

Abstract

La présente invention concerne un procédé, un dispositif et un appareil de recommandation de vidéos et un support de stockage. Le procédé comporte les étapes consistant à: acquérir un ensemble de vidéos actuelles; introduire l'ensemble de vidéos actuelles et un ensemble de vidéos historiques dans un modèle de recommandation de vidéo historiques les plus récentes pour obtenir une étiquette de catégorie et un score de catégorie pour chaque vidéo de l'ensemble de vidéos actuelles; et obtenir des vidéos recommandées sous chaque étiquette de catégorie selon les scores de catégories des vidéos sous chaque étiquette de catégorie dans l'ensemble de vidéos actuelles.
PCT/CN2019/124582 2018-12-29 2019-12-11 Procédé, dispositif et appareil de recommandation de vidéos et support de stockage WO2020135054A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811640356.6A CN111385659B (zh) 2018-12-29 2018-12-29 一种视频推荐方法、装置、设备及存储介质
CN201811640356.6 2018-12-29

Publications (1)

Publication Number Publication Date
WO2020135054A1 true WO2020135054A1 (fr) 2020-07-02

Family

ID=71128373

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/124582 WO2020135054A1 (fr) 2018-12-29 2019-12-11 Procédé, dispositif et appareil de recommandation de vidéos et support de stockage

Country Status (2)

Country Link
CN (1) CN111385659B (fr)
WO (1) WO2020135054A1 (fr)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112016613A (zh) * 2020-08-26 2020-12-01 广州市百果园信息技术有限公司 视频内容分类模型的训练方法、装置、计算机设备及介质
CN112364202B (zh) * 2020-11-06 2023-11-14 上海众源网络有限公司 视频推荐方法、装置及电子设备

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20120049550A (ko) * 2010-11-09 2012-05-17 경희대학교 산학협력단 시간에 따라 변화하는 사용자 선호도를 고려한 컨텐츠 추천 방법
CN104065981A (zh) * 2014-06-20 2014-09-24 海信集团有限公司 一种视频推荐方法和装置
CN108228911A (zh) * 2018-02-11 2018-06-29 北京搜狐新媒体信息技术有限公司 一种相似视频的计算方法及装置
CN108243357A (zh) * 2018-01-25 2018-07-03 北京搜狐新媒体信息技术有限公司 一种视频推荐方法及装置
CN109104620A (zh) * 2018-07-26 2018-12-28 腾讯科技（深圳）有限公司 一种短视频推荐方法、装置和可读介质

Family Cites Families (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101112092A (zh) * 2005-05-09 2008-01-23 皇家飞利浦电子股份有限公司 推荐器中的辅助用户控制
CN101834837A (zh) * 2009-12-18 2010-09-15 北京邮电大学 基于宽带网络的旅游景区景点在线景观视频主动信息服务系统
CN102651033B (zh) * 2012-04-09 2016-04-27 百度在线网络技术（北京）有限公司 一种在线资源的推荐方法和装置
US20140215506A1 (en) * 2013-01-25 2014-07-31 Mobitv, Inc. Time context weighted content recommendation
US9535897B2 (en) * 2013-12-20 2017-01-03 Google Inc. Content recommendation system using a neural network language model
CN103838835B (zh) * 2014-02-25 2017-11-21 中国科学院自动化研究所 一种网络敏感视频检测方法
US10460247B2 (en) * 2015-12-08 2019-10-29 Adobe Inc. Attribute weighting for media content-based recommendation
CN105512331B (zh) * 2015-12-28 2019-03-26 海信集团有限公司 一种视频推荐方法及装置
CN107451148A (zh) * 2016-05-31 2017-12-08 北京金山安全软件有限公司 一种视频分类方法、装置及电子设备
CN106294783A (zh) * 2016-08-12 2017-01-04 乐视控股（北京）有限公司 一种视频推荐方法及装置
KR102012676B1 (ko) * 2016-10-19 2019-08-21 삼성에스디에스 주식회사 콘텐츠 추천 방법, 장치 및 시스템
CN106407477A (zh) * 2016-11-22 2017-02-15 深圳市互联在线云计算股份有限公司 一种多维互联推荐方法及系统
JP2018116515A (ja) * 2017-01-19 2018-07-26 株式会社日立製作所 制御装置
CN107194419A (zh) * 2017-05-10 2017-09-22 百度在线网络技术（北京）有限公司 视频分类方法及装置、计算机设备与可读介质
CN107515909B (zh) * 2017-08-11 2020-05-19 深圳市云网拜特科技有限公司 一种视频推荐方法及系统
CN107562875A (zh) * 2017-08-31 2018-01-09 北京麒麟合盛网络技术有限公司 一种模型的更新方法、装置及系统
CN107846629B (zh) * 2017-10-11 2021-01-26 五八有限公司 向用户推荐视频的方法、装置和服务器
CN107563455A (zh) * 2017-10-18 2018-01-09 百度在线网络技术（北京）有限公司 用于获取信息的方法及装置
CN107743249A (zh) * 2017-11-27 2018-02-27 四川长虹电器股份有限公司 一种基于模型融合的ctr预估方法
CN108647293B (zh) * 2018-05-07 2022-02-01 广州虎牙信息科技有限公司 视频推荐方法、装置、存储介质和服务器

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20120049550A (ko) * 2010-11-09 2012-05-17 경희대학교 산학협력단 시간에 따라 변화하는 사용자 선호도를 고려한 컨텐츠 추천 방법
CN104065981A (zh) * 2014-06-20 2014-09-24 海信集团有限公司 一种视频推荐方法和装置
CN108243357A (zh) * 2018-01-25 2018-07-03 北京搜狐新媒体信息技术有限公司 一种视频推荐方法及装置
CN108228911A (zh) * 2018-02-11 2018-06-29 北京搜狐新媒体信息技术有限公司 一种相似视频的计算方法及装置
CN109104620A (zh) * 2018-07-26 2018-12-28 腾讯科技（深圳）有限公司 一种短视频推荐方法、装置和可读介质

Also Published As

Publication number Publication date
CN111385659B (zh) 2021-08-17
CN111385659A (zh) 2020-07-07

Similar Documents

Publication Publication Date Title
Yuan et al. Active learning for deep visual tracking
CN107491432B (zh) 基于人工智能的低质量文章识别方法及装置、设备及介质
CN110069709B (zh) 意图识别方法、装置、计算机可读介质及电子设备
CN107766929B (zh) 模型分析方法及装置
Murray et al. A deep architecture for unified aesthetic prediction
WO2020253127A1 (fr) Procédé et appareil d'apprentissage de modèle d'extraction de caractéristiques faciales, procédé et appareil d'extraction de caractéristiques faciales, dispositif et support d'informations
CN104573130B (zh) 基于群体计算的实体解析方法及装置
CN103559504A (zh) 图像目标类别识别方法及装置
Hua et al. Collaborative active visual recognition from crowds: A distributed ensemble approach
WO2020082734A1 (fr) Procédé et appareil de reconnaissance d'émotions dans un texte, dispositif électronique, et support d'enregistrement non volatil lisible par ordinateur
CN110880019A (zh) 通过无监督域适应训练目标域分类模型的方法
CN112749330B (zh) 信息推送方法、装置、计算机设备和存储介质
Xu et al. Training robust object detectors from noisy category labels and imprecise bounding boxes
CN110807086A (zh) 文本数据标注方法及装置、存储介质、电子设备
CN109214444B (zh) 基于孪生神经网络和gmm的游戏防沉迷判定系统及方法
CN111695458A (zh) 一种视频图像帧处理方法及装置
WO2023088174A1 (fr) Procédé et appareil de détection de cible
WO2020135054A1 (fr) Procédé, dispositif et appareil de recommandation de vidéos et support de stockage
CN110458600A (zh) 画像模型训练方法、装置、计算机设备及存储介质
Gu et al. Unsupervised and semi-supervised robust spherical space domain adaptation
CN113239159A (zh) 基于关系推理网络的视频和文本的跨模态检索方法
Jiang et al. Boosting facial expression recognition by a semi-supervised progressive teacher
Yu et al. Deep metric learning with dynamic margin hard sampling loss for face verification
CN112711676B (zh) 一种视频召回方法、装置、电子设备及存储介质
CN113657473A (zh) 一种基于迁移学习的Web服务分类方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19903396

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19903396

Country of ref document: EP

Kind code of ref document: A1