Content recommendation method, content recommendation model training method and device

Abstract

The present disclosure relates to a content recommendation method, a training method and device of a content recommendation model. The method comprises: obtaining candidate content associated with trigger content, and object information of a first type of object that has interacted with the trigger content; obtaining, through a first interaction prediction model corresponding to a first display mode, first content feature information of the candidate content in the first display mode, and first object feature information of the first type of object in the first display mode; determining second interaction prediction information of a to-be-recommended object to the candidate content in a second display mode based on content feature information of the trigger content, the first content feature information, and the first object feature information; and determining, based on the second interaction prediction information, a recommended content from the candidate content, and pushing the recommended content to the to-be-recommended object in the second display mode. The method can ensure the diversity and interest of the form of content watched by the user.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer, and particularly relates to a content recommendation method and device, electronic equipment, storage medium and computer program product. BACKGROUND

[0002] With the development of new media, watching content such as short videos on the platform of new media has become a main way of daily leisure and entertainment for people. At present, on the platform of new media, after a user actively selects a content for watching, some content related to the content for watching will be recommended for the user to watch.

[0003] In the current content recommendation method, content recommendation is usually performed in a double-column active selection display mode or in a passive display mode of sliding down. The application mode of the recommendation method is single. However, the two display modes have advantages and disadvantages. Therefore, how to better combine the two display modes to recommend content to users is the key to improving the browsing experience of users. SUMMARY

[0004] The present disclosure provides a content recommendation method, device, electronic equipment and storage medium to at least solve the problem of single application mode of the content recommendation method in the related art. The technical solutions of the present disclosure are as follows.

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

[0006] obtaining candidate content associated with trigger content and object information of a first type of object that has interacted with the trigger content; the first type of object is an object that performs an interaction behavior in a first display mode, and the first display mode is a content display mode before the trigger content is triggered;

[0007] obtaining, by a first interaction prediction model corresponding to the first display mode, first content feature information of the candidate content in the first display mode and first object feature information of the first type of object in the first display mode; the first interaction prediction model is used to determine first interaction prediction information of a to-be-recommended content relative to the first type of object in the first display mode;

[0008] determining, based on content feature information of the trigger content, the first content feature information and the first object feature information, second interaction prediction information of a to-be-recommended object to the candidate content in a second display mode; the second display mode is a content display mode entered after the trigger content is triggered;

[0009] determine recommended content from the candidate content based on the second interaction prediction information, and push the recommended content to the to-be-recommended object in the second display mode.

[0010] In an example embodiment, the determining of the second interaction prediction information of the to-be-recommended object to the candidate content in the second display mode based on the content feature information of the trigger content, the first content feature information, and the first object feature information comprises:

[0011] obtaining at least one of the historical interaction information of the to-be-recommended object to the content in the second display mode, the content feature information of the candidate content, and the object feature information of a second type of object that has interacted with the candidate content; the second type of object is an object that performs an interaction behavior in the second display mode;

[0012] determining the second interaction prediction information of the to-be-recommended object to the candidate content in the second display mode based on the at least one information, and the content feature information of the trigger content, the first content feature information, and the first object feature information.

[0013] In an example embodiment, the determining of the second interaction prediction information of the to-be-recommended object to the candidate content in the second display mode based on the at least one information, and the content feature information of the trigger content, the first content feature information, and the first object feature information comprises:

[0014] when the at least one information includes the historical interaction information and the content feature information of the candidate content, performing feature extraction processing on the historical interaction information and the content feature information of the candidate content through an attention unit to obtain extracted feature information;

[0015] obtaining the second interaction prediction information of the to-be-recommended object to the candidate content based on the extracted feature information, the content feature information of the candidate content, and the content feature information of the trigger content, the first content feature information, and the object feature information of the first type of object.

[0016] In an example embodiment, the determining of the recommended content from the candidate content based on the second interaction prediction information comprises:

[0017] performing queue division on the candidate content according to matching information of the candidate content and the trigger content on a plurality of preset attribute information to obtain a plurality of candidate content sequences;

[0018] performing sorting on the candidate content in each of the candidate content sequences according to the second interaction prediction information to obtain a plurality of sorted candidate content sequences.

[0019] The recommended content is determined from at least one of the sorted candidate content sequences according to a priority order of each of the candidate content sequences.

[0020] In an example embodiment, the queue division of the candidate content according to the matching information of the candidate content and the trigger content on the preset attribute information to obtain a plurality of candidate content sequences comprises:

[0021] A plurality of empty queues are obtained; the empty queues are determined based on a priority order of each of the attribute information and a matching number on the attribute information;

[0022] Each of the candidate contents is divided into each of the empty queues according to the matching information of the candidate content and the trigger content and the priority order of each of the attribute information to obtain a plurality of candidate content sequences.

[0023] In an example embodiment, the determination of the recommended content from at least one of the sorted candidate content sequences according to a priority order of each of the candidate content sequences comprises:

[0024] A target number of required recommended content is obtained;

[0025] The candidate content corresponding to the target number is selected from at least one of the sorted candidate content sequences as the recommended content in a descending order of the priority of each of the candidate content sequences.

[0026] In an example embodiment, the determination of the second interaction prediction information of the candidate content by the to-be-recommended object in the second display mode based on the content feature information of the trigger content, the first content feature information, and the first object feature information comprises:

[0027] The content feature information of the trigger content, the first content feature information, and the first object feature information are subjected to information prediction processing by a second interaction prediction model corresponding to the second display mode to obtain the second interaction prediction information of the candidate content by the to-be-recommended object in the second display mode.

[0028] According to a second aspect of the embodiments of the present disclosure, a training method of a content recommendation model is provided, comprising:

[0029] obtaining sample interaction data; the sample interaction data includes sample content feature information of a sample trigger content, first sample content feature information of a sample candidate content associated with the sample trigger content in a first display mode, second interaction information of a second type of sample object to the sample candidate content, and first sample object feature information of a first type of sample object that has interacted with the sample trigger content in the first display mode; the first sample content feature information and the first sample object feature information are obtained by a first interaction prediction model corresponding to the first display mode;

[0030] performing information prediction processing on the sample content feature information, the first sample content feature information, and the first sample object feature information by a second interaction prediction model to be trained, to obtain second interaction prediction information of a second type of sample object to the sample candidate content in a second display mode;

[0031] training the second interaction prediction model to be trained based on a loss value between the second interaction prediction information and the second interaction information, to obtain a trained second interaction prediction model as a content recommendation model;

[0032] The first display mode is a content display mode before the sample trigger content is triggered, the second display mode is a content display mode entered after the sample trigger content is triggered, the first type of sample object is an object that performs an interaction behavior in the first display mode, the second type of sample object is an object that performs an interaction behavior in the second display mode, and the first interaction prediction model is used to determine first interaction prediction information of a content to be recommended relative to the first type of object in the first display mode.

[0033] According to a third aspect of the embodiments of the present disclosure, a content recommendation device is provided, including:

[0034] The obtaining unit is configured to perform obtaining candidate content associated with trigger content and object information of a first type of object that has interacted with the trigger content; the first type of object is an object that performs an interaction behavior in a first display mode, and the first display mode is a content display mode before the trigger content is triggered;

[0035] The determining unit is configured to perform obtaining, by a first interaction prediction model corresponding to the first display mode, first content feature information of the candidate content in the first display mode and first object feature information of the first type of object in the first display mode; the first interaction prediction model is used to determine first interaction prediction information of a content to be recommended relative to the first type of object in the first display mode;

[0036] a prediction unit configured to determine second interaction prediction information of the candidate content by the to-be-recommended object in a second display mode based on the content feature information of the trigger content, the first content feature information, and the first object feature information, the second display mode being a content display mode entered after the trigger content is triggered;

[0037] a recommendation unit configured to determine a recommended content from the candidate content based on the second interaction prediction information, and push the recommended content to the to-be-recommended object in the second display mode.

[0038] In an example embodiment, the prediction unit is further configured to obtain at least one of historical interaction information of the to-be-recommended object for content in the second display mode, content feature information of the candidate content, and object feature information of a second type of object that has interacted with the candidate content, the second type of object being an object that performs an interaction behavior in the second display mode; and determine the second interaction prediction information of the candidate content by the to-be-recommended object in the second display mode based on the at least one information, the content feature information of the trigger content, the first content feature information, and the first object feature information.

[0039] In an example embodiment, the prediction unit is further configured to, when the at least one information includes the historical interaction information and the content feature information of the candidate content, perform feature extraction processing on the historical interaction information and the content feature information of the candidate content by an attention unit to obtain extracted feature information; and obtain the second interaction prediction information of the candidate content by the to-be-recommended object based on the extracted feature information, the content feature information of the candidate content, and the content feature information of the trigger content, the first content feature information, and the object feature information of the first type of object.

[0040] In an example embodiment, the recommendation unit is further configured to perform queue division on the candidate content according to matching information of the candidate content and the trigger content on a plurality of preset attribute information to obtain a plurality of candidate content sequences; perform sorting on the candidate content in each of the candidate content sequences according to the second interaction prediction information to obtain a plurality of sorted candidate content sequences; and determine the recommended content from at least one of the sorted candidate content sequences according to a priority order of each of the candidate content sequences.

[0041] In an example embodiment, the recommendation unit is further configured to perform obtaining a plurality of empty queues preset; the empty queues are determined based on a priority order of the attribute information and a matching number on the attribute information; and each of the candidate contents is divided into each of the empty queues according to the matching information corresponding to the trigger content and the priority order of each of the attribute information, to obtain a plurality of candidate content sequences.

[0042] In an example embodiment, the recommendation unit is further configured to perform obtaining a target number of required recommendation contents; and selecting candidate contents corresponding to the target number from at least one of the sorted candidate content sequences in a descending order of priority of each of the candidate content sequences, as the recommendation contents.

[0043] In an example embodiment, the prediction unit is further configured to perform information prediction processing on the content feature information of the trigger content, the first content feature information, and the first object feature information by a second interaction prediction model corresponding to the second display mode, to obtain second interaction prediction information of the to-be-recommended object on the candidate content in the second display mode.

[0044] According to a fourth aspect of the embodiments of the present disclosure, a training device of a content recommendation model is provided, including:

[0045] The obtaining unit is configured to perform obtaining sample interaction data; the sample interaction data includes sample content feature information of sample trigger content, first sample content feature information of sample candidate content associated with the sample trigger content in a first display mode, second interaction information of a second type of sample object on the sample candidate content, and first sample object feature information of a first type of sample object interacting with the sample trigger content in the first display mode; the first sample content feature information and the first sample object feature information are obtained by a first interaction prediction model corresponding to the first display mode;

[0046] The prediction unit is configured to perform information prediction processing on the sample content feature information, the first sample content feature information, and the first sample object feature information by a second interaction prediction model to be trained, to obtain second interaction prediction information of the second type of sample object on the sample candidate content in a second display mode.

[0047] The training unit is configured to perform training the second interaction prediction model to be trained based on a loss value between the second interaction prediction information and the second interaction information, to obtain a trained second interaction prediction model as a content recommendation model.

[0048] The first display mode is a content display mode before the sample triggered content is triggered, the second display mode is a content display mode entered after the sample triggered content is triggered, the first type of sample object is an object that performs an interaction behavior in the first display mode, the second type of sample object is an object that performs an interaction behavior in the second display mode, and the first interaction prediction model is used to determine first interaction prediction information of to-be-recommended content relative to the first type of object in the first display mode.

[0049] According to a fifth aspect of an embodiment of the present disclosure, an electronic device is provided, comprising:

[0050] a processor;

[0051] a memory for storing instructions executable by the processor;

[0052] The processor is configured to execute the instructions to implement the method according to any one of the preceding aspects.

[0053] According to a sixth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device can execute the method according to any one of the preceding aspects.

[0054] According to a seventh aspect of an embodiment of the present disclosure, a computer program product is provided, the computer program product comprises instructions, when the instructions are executed by a processor of an electronic device, the electronic device can execute the method according to any one of the preceding aspects.

[0055] The technical solutions provided by the embodiments of the present disclosure at least bring the following beneficial effects:

[0056] After obtaining the candidate content associated with the trigger content and the object information of the first type of object that has interacted with the trigger content, the first content feature information of the candidate content in the first display mode and the first object feature information of the first type of object in the first display mode are obtained through a first interaction prediction model corresponding to the first display mode. Further, based on the content feature information of the trigger content, the first content feature information, and the first object feature information, the second interaction prediction information of the candidate content by the to-be-recommended object in the second display mode is determined. Based on the second interaction prediction information, the recommended content is determined from the candidate content, and the recommended content is pushed to the to-be-recommended object in the second display mode. This method increases the second display mode in the detail page of the trigger content selected by the to-be-recommended object, and determines the second interaction prediction information of the candidate content by the to-be-recommended object in the second display mode based on the first content feature information and the first object feature information of the candidate content and the first type of object in the first display mode, and selects the recommended content based on the first interaction prediction information, fully considers the linkage relationship between the trigger content and the second display mode, realizes the smooth transition between the two display modes, can guarantee the diversity and interest of the content form watched by the user, retains the right of active selection and provides passive selection, and there is a certain correlation between the two, thereby improving the consumption experience and consumption time of the user.

[0057] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0058] The accompanying drawings incorporated in the specification and forming a part of it illustrate embodiments consistent with the present disclosure and together with the description serve to explain the principles of the present disclosure, and do not constitute an undue limitation on the disclosure.

[0059] Figure 1 is a flowchart of a content recommendation method according to an exemplary embodiment.

[0060] Figure 2 is a model structure diagram of an interaction prediction model according to an exemplary embodiment.

[0061] Figure 3 is a flowchart of a recommended content determination step according to an exemplary embodiment.

[0062] Figure 4 is a flowchart of a training method of a content recommendation model according to an exemplary embodiment.

[0063] Figure 5 is a structural block diagram of a content recommendation device according to an exemplary embodiment.

[0064] Figure 6 is a structural block diagram of a content recommendation model training device according to an exemplary embodiment.

[0065] Figure 7 is a block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0066] In order for those skilled 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 described clearly and completely below with reference to the drawings.

[0067] It should be noted that the terms "first", "second" and the like in the specification and claims of the present disclosure and the above-described drawings are used to distinguish similar objects, and do not necessarily have to describe a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than that illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all the embodiments consistent with the present disclosure. Rather, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims. It should also be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, analyzed data, etc.) involved in the present disclosure are all information and data authorized by the user or authorized by all parties.

[0068] It should be noted that the first display mode in the present disclosure is the content display mode before triggering the content trigger, and the second display mode is the content display mode entered after triggering the content trigger. Specifically, in one embodiment, the first display mode can be a multi-column active selection display mode, and the second display mode can be a passive display mode of sliding down, for example, the first display mode is a double waterfall display mode, and the second display mode is a single waterfall display mode, or the first display mode is a three-column waterfall display mode, and the second display mode is a single waterfall display mode, etc. In another embodiment, the first display mode and the second display mode can also be both multi-column active selection display modes, but the number of columns of the first display mode and the second display mode is different. For example, the first display mode is a three-column waterfall display mode, and the second display mode is a double waterfall display mode, or the first display mode is a double waterfall display mode, and the second display mode is a three-column waterfall display mode. In actual application, those skilled in the art can set the specific forms of the first display mode and the second display mode according to the needs, and the present disclosure does not make specific limitations thereon. For the sake of illustration, the embodiments of the present disclosure will be described below taking the double-column active selection display mode as the first display mode and the single-column passive display mode of sliding down as the second display mode.

[0069] In an example embodiment, as shown in Figure 1 A content recommendation method is provided, and the embodiment is exemplified by the method applied to a terminal. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction of the terminal and the server. The terminal can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The Internet of Things device can be a smart speaker, a smart television, a smart air conditioner, a smart vehicle device, etc. The portable wearable device can be a smart watch, a smart bracelet, a head-mounted device, etc. The server can be implemented by an independent server or a server cluster composed of multiple servers.

[0070] The method includes the following steps in the embodiment:

[0071] In step S110, candidate content associated with the trigger content and object information of a first type of object that has interacted with the trigger content are obtained. The first type of object is an object that performs an interaction behavior in a first display mode, and the first display mode is a content display mode before the trigger content is triggered.

[0072] The trigger content can be understood as content that a to-be-recommended object actively selects to browse, and the form of the trigger content can be a video, text, or an image, etc.

[0073] The object information can be understood as information representing the identity, interests, and other characteristics of the object, for example, the object information can be age, occupation, interests, and other information.

[0074] The first display mode can be a double-column active selection display mode.

[0075] In a specific implementation, the trigger content can be displayed on a main page in an application program, and the content on the main page is displayed in the first display mode. The user can select any content on the main page for browsing and viewing through a sliding operation. The content selected and triggered by the to-be-recommended object on the main page is the trigger content. In actual application, a plurality of contents associated with each content on the main page can be pre-configured as candidate contents. When a triggering operation of the to-be-recommended object on any content on the main page is received, the triggered content is taken as the trigger content, and the candidate content associated with the trigger content is obtained. At the same time, the object information of the first type of object that has interacted with the trigger content is also obtained. Specifically, the object information of the first type of object that has performed at least one interaction behavior such as liking, commenting, collecting, watching, and forwarding the trigger content can be obtained.

[0076] For example, in a short video watching platform, on a video home page displayed in a double-column active selection display mode, a user can select any video in the video home page for watching through a sliding operation. The video triggered in the home page is a trigger video. Candidate videos associated with each trigger video can be pre-configured to facilitate subsequent selection of a recommended video from the candidate videos and recommendation to the user. After the user selects the trigger video, the terminal can display the recommended video associated with the trigger video for the user to watch.

[0077] In step S120, first content feature information of the candidate content in the first display mode and first object feature information of the first type of object in the first display mode are obtained through a first interaction prediction model corresponding to the first display mode. The first interaction prediction model is used to determine first interaction prediction information of the content to be recommended relative to the first type of object in the first display mode.

[0078] The first interaction prediction model can be used to obtain features of any content or any object in the first display mode, and obtain first interaction prediction information of any object and any content based on the features.

[0079] In a specific implementation, sample content in the first display mode and a first type of sample object can be pre-obtained, and the first interaction prediction model to be trained is trained through the sample content and the first type of sample object to obtain a trained first interaction prediction model. More specifically, refer to Figure 2 The model below the dotted line is a model structure diagram of the first interaction prediction model, which includes two multi-layer perceptrons for feature extraction of the sample content and the first type of sample object in the first display mode to obtain first sample content feature information of the sample content and first sample object feature information of the first type of sample object. The first sample content feature information and the first sample object feature information are further dot multiplied to obtain a click rate. The first interaction prediction model corresponding to the first display mode is trained through the click rate. The content feature information of the candidate content associated with the trigger content can be further extracted through the first interaction prediction model to obtain first content feature information of the candidate content in the first display mode. The object information of the first type of object is extracted through the first interaction prediction model to obtain first object feature information of the first type of object in the first display mode.

[0080] In step S130, based on the content feature information of the trigger content, the first content feature information, and the first object feature information, second interaction prediction information of the candidate content by the object to be recommended in the second display mode is determined. The second display mode is a content display mode entered after the trigger content is triggered.

[0081] The content feature information is information representing a content category and a label, and can be understood as description information of the content.

[0082] The second display mode can be a downward passive display mode.

[0083] The second interaction prediction information can be a viewing duration or other information representing an interaction condition.

[0084] In a specific implementation, after obtaining the content feature information of the trigger content, the first content feature information of the candidate content in the first display mode, and the first object feature information of the first type of object in the first display mode, the content feature information of the trigger content, the first content feature information, and the first object feature information can be subjected to information prediction processing by a second interaction prediction model corresponding to the second display mode, to obtain second interaction prediction information of the to-be-recommended object to the candidate content in the second display mode. More specifically, the content feature information of the trigger content can be subjected to feature extraction processing by a feature extraction unit, to obtain extracted feature information of the trigger content. Based on the extracted feature information, the first content feature information, and the first object feature information, the second interaction prediction information of the to-be-recommended object to the candidate content in the second display mode can be determined.

[0085] In step S140, the second interaction prediction information is used to determine a recommended content from the candidate content, and the recommended content is pushed to the to-be-recommended object in the second display mode.

[0086] In a specific implementation, at least one recommended content can be determined from the candidate content in descending order of interest degree represented by the second interaction prediction information, and the determined at least one recommended content can be pushed to the to-be-recommended object in the second display mode.

[0087] For example, taking the second interaction prediction information as a viewing duration as an example, the longer the viewing duration, the higher the interest degree. Therefore, the candidate content can be sorted in descending order of viewing duration, and a recommended content with a viewing duration in the top N (N≥1) can be determined from the sorted candidate content and pushed to the to-be-recommended object.

[0088] In the content recommendation method, after obtaining the candidate content associated with the trigger content and the object information of the first type of object that has interacted with the trigger content, the first content feature information of the candidate content in the first display mode and the first object feature information of the first type of object in the first display mode are obtained through the first interaction prediction model corresponding to the first display mode. Further, the second interaction prediction information of the candidate content by the to-be-recommended object in the second display mode is determined based on the content feature information of the trigger content, the first content feature information, and the first object feature information. The recommendation content is determined from the candidate content based on the second interaction prediction information, and the recommendation content is pushed to the to-be-recommended object in the second display mode. This method increases the second display mode in the detail page of the trigger content selected by the to-be-recommended object, and determines the second interaction prediction information of the candidate content by the to-be-recommended object in the second display mode based on the first content feature information and the first object feature information of the candidate content and the first type of object in the first display mode, and selects the recommendation content based on the first interaction prediction information. The linkage relationship between the trigger content and the second display mode is fully considered, the smooth transition between the two display modes is realized, the diversity and interest of the content viewing form of the user are guaranteed, the right of active selection is reserved and passive selection is provided, and there is a certain correlation between the two, thereby improving the consumption experience and consumption time of the user.

[0089] To further improve the accuracy of the determined second interaction prediction information of the candidate content by the to-be-recommended object in the second display mode, the present disclosure further provides another exemplary embodiment, in which the above step S130 can be implemented by the following steps:

[0090] Step S130A, obtaining at least one of the historical interaction information of the to-be-recommended object for the content in the second display mode, the content feature information of the candidate content, and the object feature information of the second type of object that has interacted with the candidate content; the second type of object is an object that performs an interaction behavior in the second display mode;

[0091] Step S130B, determining the second interaction prediction information of the candidate content by the to-be-recommended object in the second display mode based on the at least one information, and the content feature information of the trigger content, the first content feature information, and the first object feature information.

[0092] In a specific implementation, reference is made to Figure 2The model structure diagram of the second interaction prediction model in the upper side model in FIG. 11 is shown in FIG. 12. As shown in the figure, when determining the second interaction prediction information of the to-be-recommended object to the candidate content in the second display mode, the input conditions can include at least one of the historical interaction information of the to-be-recommended object to the content in the second display mode, the content feature information of the candidate content, and the object feature information of the second type of object that has interacted with the candidate content, in addition to the content feature information of the trigger content, the first content feature information of the candidate content in the first display mode, and the first object feature information of the first type of object in the first display mode. Through the second interaction prediction model, the at least one information, the content feature information of the trigger content, the first content feature information, and the first object feature information are subjected to information prediction processing to obtain the second interaction prediction information of the to-be-recommended object to the candidate content in the second display mode.

[0093] In the embodiment, on the basis of the content feature information of the trigger content, the first content feature information, and the first object feature information, at least one of the historical interaction information of the to-be-recommended object to the content in the second display mode, the content feature information of the candidate content, and the object feature information of the second type of object that has interacted with the candidate content is combined to implement the prediction of the second interaction prediction information from multiple dimensions, which can improve the accuracy of the determined second interaction prediction information.

[0094] Further, in an exemplary embodiment, the above step S103B can be implemented in the following manner:

[0095] In step S130B-1, when the at least one information includes the historical interaction information and the content feature information of the candidate content, the historical interaction information and the content feature information of the candidate content are subjected to feature extraction processing through the attention unit to obtain the extracted feature information.

[0096] In step S130B-2, based on the extracted feature information, the content feature information of the candidate content, and the content feature information of the trigger content, the first content feature information, and the object feature information of the first type of object, the second interaction prediction information of the to-be-recommended object to the candidate content is obtained.

[0097] In a specific implementation, reference can be made to Figure 2As shown in the model structure diagram of the second interaction prediction model, when the at least one information includes the historical interaction information and the content feature information of the candidate content, the historical interaction information and the content feature information of the candidate content can be subjected to feature extraction processing through an attention unit (Attention mechanism), to obtain extracted feature information. The extracted feature information, the content feature information of the candidate content, and the content feature information of the trigger content, the first content feature information, and the object feature information of the first type of object are subjected to merging processing, to obtain second interaction prediction information of the to-be-recommended object on the candidate content.

[0098] More specifically, the historical interaction information and the content feature information of the candidate content are subjected to feature extraction processing through the attention unit, a weight distribution of the historical interaction information is learned through the attention unit, the weight distribution is applied to the content feature information of the candidate content, and thus important feature information of the candidate content can be selected, and irrelevant feature information can be inhibited.

[0099] In this embodiment, the historical interaction information and the content feature information of the candidate content are subjected to feature extraction processing through the attention unit, a weight distribution of the interest of the to-be-recommended object can be learned based on the historical interaction information, and thus content feature information of the candidate content that is more consistent with the to-be-recommended object can be extracted, the determination of the subsequent second interaction prediction information is performed, and the adaptation degree of the determined second interaction prediction information to the to-be-recommended object is improved.

[0100] In an exemplary embodiment, as shown in Figure 3 The step S140 can be implemented by the following steps.

[0101] In step S310, the candidate contents are subjected to queue division according to matching information of the candidate contents and the trigger content on a plurality of preset attribute information, to obtain a plurality of candidate content sequences.

[0102] The attribute information represents expression attributes of the content itself, and can include identification information (also referred to as ID information), classification information, and label information.

[0103] The matching information represents matching results of the candidate content and the trigger content on each attribute information, for example, the matching information can be that the identification information of the candidate content and the trigger content is matched, the classification information is not matched, and the label information is matched.

[0104] In a specific implementation, after determining the candidate content associated with the trigger content, the identification information, classification information and label information of the candidate content and the trigger content are respectively obtained, the respective attribute information of the candidate content and the trigger content is matched respectively, the matching information of the candidate content and the trigger content on the attribute information such as the identification information, the classification information and the label information is obtained, and the candidate content is further divided into the plurality of empty queues according to the matching information, so as to obtain the plurality of candidate content sequences.

[0105] For example, the identification information of the trigger content is A, the classification information is B, and the label information is C, the identification information of the candidate content is A * , the classification information is B * , and the label information is C * . The identification information A is matched with the identification information A * , the classification information B is matched with the classification information B * , and the label information C is matched with the label information C * , so as to obtain the identification information matching result, the classification information matching result and the label information matching result of the candidate content and the trigger content, and the identification information matching result, the classification information matching result and the label information matching result are taken as the matching information of the candidate content and the trigger content.

[0106] In step S320, the candidate content in each candidate content sequence is sorted according to the second interaction prediction information, so as to obtain the plurality of sorted candidate content sequences.

[0107] In a specific implementation, the candidate content in each candidate content sequence is sorted in the order from large to small of the interest degree represented by the second interaction prediction information, so as to obtain the plurality of sorted candidate content sequences.

[0108] In step S330, the recommended content is determined from at least one sorted content sequence according to the priority order of each candidate content sequence.

[0109] The priority order of each candidate content sequence can be determined based on the priority order of each attribute information and the number of matching attribute information. The more the number of matching attribute information, the higher the priority of the matching attribute information, and the higher the priority. The priority order of each attribute information can be determined based on the fineness of the granularity of the attribute information. The finer the granularity, the higher the priority. For example, the priority order of each attribute information can be: identification information > classification information > label information.

[0110] In a specific implementation, the candidate content sequences can be sorted according to the priority order of the candidate content sequences, and the selection of the recommended content can be performed from the candidate content sequence with the highest priority first. When the candidate content in the candidate content sequence with the highest priority is selected, the selection of the recommended content can be continued from the candidate content sequence with the second highest priority, and so on. When the selection of the recommended content is performed in each sorted content sequence, the selection of the recommended content can be performed according to the order of the interest degree from large to small represented by the second interaction prediction information.

[0111] In the embodiment, after the candidate content associated with the trigger content is obtained, the candidate content is divided into queues according to the matching information of the candidate content and the trigger content on the preset multiple attribute information, to obtain multiple candidate content sequences, which realizes the sorting of the candidate content in the attribute information dimension. The candidate content in each candidate content sequence is sorted according to the second interaction prediction information of the to-be-recommended object on the candidate content, so that each candidate content sequence is also ordered internally. Finally, the recommended content is selected from at least one sorted candidate content sequence according to the priority order of each candidate content sequence, so that the content more related to the trigger content is more likely to be retained as the recommended content, thereby improving the relevance of the recommended content and the trigger content and improving the adaptability of the recommended content to the to-be-recommended object.

[0112] In an exemplary embodiment, in step S310, the candidate content is divided into queues according to the matching information of the candidate content and the trigger content on the preset multiple attribute information, to obtain multiple candidate content sequences, which can be implemented by the following steps:

[0113] In step S310A, multiple empty queues are obtained. The empty queues are determined based on the priority order of each attribute information and the matching number on the attribute information.

[0114] In step S310B, each candidate content is divided into each empty queue according to the matching information corresponding to the candidate content and the trigger content and the priority order of each attribute information, to obtain multiple candidate content sequences.

[0115] In a specific implementation, since the attribute information has multiple attribute information, the matching number of the candidate content and the trigger content on the attribute information also has multiple matching numbers. Therefore, multiple empty queues can be preset according to the matching number on the attribute information. Under the same matching number, further queue subdivision can be performed according to the priority order of each attribute information.

[0116] For example, if there are three attribute information, such as identification information, classification information and label information, and the priority order is: identification information > classification information > label information, then the matching number can be divided into three empty queues, i.e., matching number 3, matching number 2 and matching number 1. Further, when the matching number is 2 and 1, the queues can be further subdivided according to the priority order of the attribute information. For example, when the matching number is 2, the queues can be further subdivided into three queues, i.e., identification information and classification information both match, identification information and label information both match, and classification information and label information both match. When the matching number is 1, the queues can be further subdivided into three queues, i.e., identification information matches, classification information matches, and label information matches. Thus, when there are three attribute information, 1+3+3=7 empty queues can be preset.

[0117] In the preset empty queues, the candidate content can be divided into the empty queues according to the matching information corresponding to the candidate content and the trigger content, and the priority order of the attribute information, to obtain a plurality of candidate content sequences.

[0118] For example, if the candidate content and the trigger content both match the identification information and the label information, and the corresponding matching number is 2, then the candidate content can be divided into the queue in which the identification information and the label information both match.

[0119] In this embodiment, on the one hand, the initial empty queues are divided according to the matching number on the attribute information, and on the other hand, when the matching number is the same, the queues are further subdivided according to the priority order of the attribute information, so that the queues are divided from two dimensions of the matching number and the attribute information, and the accuracy of the candidate content division result can be ensured.

[0120] In an exemplary embodiment, in the step S330, the recommended content is determined from at least one ordered candidate content sequence according to the priority order of each candidate content sequence, which can be achieved by the following steps:

[0121] In step S330A, the target number of the required recommended content is obtained.

[0122] In step S330B, the candidate content corresponding to the target number is selected from at least one ordered candidate content sequence as the recommended content according to the priority order from high to low of each candidate content sequence.

[0123] In a specific implementation, when the total number of candidate contents in the candidate content sequence with the highest priority is greater than or equal to the target number, candidate contents corresponding to the target number are selected from the candidate content sequence with the highest priority in descending order of the interest degree represented by the second interaction prediction information, as the recommended contents; when the total number of candidate contents in the candidate content sequence with the highest priority is less than the target number, candidate contents are sequentially selected from the candidate content sequence with the highest priority to the candidate content sequence with the lowest priority until the candidate contents satisfying the target number are obtained, as the recommended contents.

[0124] It can be understood that in another implementation, truncation selection can also be performed on each sorted candidate content sequence, that is, TopM candidate contents are selected from each sorted candidate content sequence according to the target number, as the recommended contents. The target number can also be divided into multiple sub-numbers in a gradient manner, and the corresponding number of candidate contents is selected from each sorted candidate content sequence, as the recommended contents. The specific selection manner of the recommended contents can be determined according to actual needs, and the present disclosure does not make specific limitations thereto.

[0125] In the embodiment, when the recommended contents are selected from each sorted candidate content sequence based on the target number of required recommended contents, on the one hand, the candidate content sequences to be selected are sequentially determined according to the priority order of each attribute information, and on the other hand, the recommended contents are selected according to the second interaction prediction information within the sequence, so that the recommended contents obtained can take into account the priority of the attribute information and the priority of the second interaction prediction information, and improve the satisfaction of the user.

[0126] In an example embodiment, in the step S130, the second interaction prediction information of the candidate content of the to-be-recommended object in the second display mode is determined based on the content feature information of the trigger content, the first content feature information, and the first object feature information, which can be achieved by the following manner: the second interaction prediction model corresponding to the second display mode is used to perform information prediction processing on the content feature information of the trigger content, the first content feature information, and the first object feature information, to obtain the second interaction prediction information of the candidate content of the to-be-recommended object in the second display mode.

[0127] In a specific implementation, the sample content feature information of the sample trigger content, the first sample content feature information of the sample candidate content associated with the sample trigger content in the first display mode, and the first sample object feature information of the first type of sample object that has interacted with the sample trigger content in the first display mode can be obtained in advance. The second interaction prediction model to be trained is trained by taking these information as input and taking the second interaction prediction information as output, to obtain the trained second interaction prediction model. Further, after obtaining the content feature information of the trigger content, the first content expression information, and the first object feature information, the trained second interaction prediction model can be used to perform information prediction processing on the content feature information of the trigger content, the first content expression information, and the first object feature information, to obtain the second interaction prediction information of the to-be-recommended object to the candidate content in the second display mode.

[0128] Referring to Figure 2 The model structure diagram of the second interaction prediction model is shown in FIG. 2. When the content feature information of the trigger content, the first content feature information, and the first object feature information are input into the second interaction prediction model, the content feature information of the trigger content can be first extracted by the multi-layer perceptron in the second interaction prediction model, to obtain the extracted feature information of the trigger content. The second interaction prediction information of the to-be-recommended object to the candidate content can be obtained by inputting the merged information into the multi-layer perceptron after merging the extracted feature information, the first content feature information, and the first object feature information.

[0129] In the embodiment, when the second interaction prediction information of the to-be-recommended object to the candidate content is estimated by the second interaction prediction model, the linkage relationship between the trigger content and the second display mode is considered, so that the accuracy of the estimated second interaction prediction information of the to-be-recommended object to the candidate content can be ensured.

[0130] In an example embodiment, as Figure 4 shown in FIG. 1, a method for training a content recommendation model is provided. In this embodiment, the method is exemplified by being applied to a terminal. It should be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and can be implemented through the interaction of the terminal and the server. In this embodiment, the method includes the following steps:

[0131] In step S410, sample interaction data is obtained; the sample interaction data includes sample content feature information of a sample trigger content, first sample content feature information of a sample candidate content associated with the sample trigger content in a first display mode, second interaction information of a second type of sample object to the sample candidate content, and first sample object feature information of a first type of sample object that has interacted with the sample trigger content in the first display mode; the first sample content feature information and the first sample object feature information are predicted by a first interaction prediction model corresponding to the first display mode;

[0132] In step S420, the sample content feature information, the first sample content feature information, and the first sample object feature information are subjected to information prediction processing by a second interaction prediction model to be trained, to obtain second interaction prediction information of the second type of sample object to the sample candidate content in a second display mode;

[0133] In step S430, the second interaction prediction model to be trained is trained based on a loss value between the second interaction prediction information and the second interaction information, to obtain a trained second interaction prediction model as a content recommendation model;

[0134] The first display mode is a content display mode before the sample trigger content triggers, the second display mode is a content display mode entered after the sample trigger content triggers, the first type of sample object is an object that performs an interaction behavior in the first display mode, the second type of sample object is an object that performs an interaction behavior in the second display mode, and the first interaction prediction model is used to determine first interaction prediction information of a to-be-recommended content relative to the first type of object in the first display mode.

[0135] In a specific implementation, the sample candidate content associated with the sample trigger content and object information of the first type of sample object that has interacted with the sample trigger content can be obtained first. The sample candidate content and the object information of the first type of object are processed by a first interaction prediction model corresponding to the first display mode, to obtain first sample content feature information of the sample candidate content in the first display mode and first object feature information of the first type of sample object in the first display mode. Second interaction information of the second type of sample object to the sample candidate content and sample content feature information of the sample trigger content are obtained, and the second interaction information, the sample content feature information, the first sample content feature information, and the first sample object feature information are combined to form sample data.

[0136] Reference Figure 2A model structure diagram of the second interaction prediction model is shown. The sample content feature information, the first sample content feature information and the first sample object feature information in the sample data are taken as model inputs to predict the second interaction prediction information of the second type of sample object to the sample candidate content as output. Regression training is performed through a loss value between the second interaction prediction information and the actual second interaction information in the sample data until a preset training number is reached or a preset accuracy is reached. Training is ended to obtain a trained second interaction prediction model as a content recommendation model.

[0137] In the embodiment, in the input information of the content recommendation model, in addition to introducing the sample content feature information of the sample trigger content, the feature information of the sample candidate content and the first type of sample object in the first display mode is combined, the linkage relationship between the selected trigger video in the first display mode and the second display mode is fully considered, the content recommendation model can capture the implicit correlation between the sample candidate content and the sample trigger content, through such model design, data migration and correlation consideration in the related business scenario are realized, the estimation result of the obtained content recommendation model is more accurate, and combination of different scenarios can be realized.

[0138] In an exemplary embodiment, in order for those skilled in the art to understand the embodiments of the present disclosure, the content recommendation method provided by the present disclosure will be described below taking the application scenario of short videos as an example.

[0139] At present, the consumption scenarios of short videos generally include an active selection consumption form in the form of double waterfall flow and a passive consumption in the form of single waterfall flow. The active selection consumption form can provide the user with the selection right, but the user operation complexity is high and the immersion is weak. The passive consumption can provide the user with a smooth viewing experience and an immersive experience. Therefore, how to better combine the two consumption forms is the key to improving the user browsing experience.

[0140] In order to be compatible with the two consumption preferences, the present scheme considers adding a passive consumption function in the video detail page after the user actively selects. However, in order to smoothly transition between the two consumptions, the present disclosure fully considers the linkage relationship between the trigger entry selected by the user and the passive consumption in the technical implementation of the recommendation system, guarantees the diversity and interest of the user consumption form, retains the right of active selection and provides passive selection, and there is a certain correlation between the two, thereby improving the user consumption experience and consumption time.

[0141] The present embodiment is applied to the linkage of the homepage waterfall flow recommendation and the video detail page recommendation flow to realize the recommendation mechanism of different flows inside and outside. Through the revision of the consumption form and the construction of the double recommendation system, and in the detail page recommendation funnel, the trigger video actively selected by the user is fully considered to complete the construction of the recommendation system with correlation and diversity. The present embodiment includes the following steps:

[0142] (1) Obtain candidate videos associated with the trigger video and obtain the user's estimated viewing time for the candidate videos.

[0143] Specifically, the estimated viewing time can be obtained through... Figure 2 The interactive prediction model shown is determined as follows: Figure 2 As shown, the interaction prediction model has a dual-tower structure. The model below the dashed line in the figure can be referred to as the first interaction prediction model, and the model above the dashed line can be referred to as the second interaction prediction model. The main goal of the entire model is to predict the watch time distribution (WTD) that users may consume on the current candidate video. The calculation method for this goal is to use 7 days of data that triggers the consumption on the video details page to statistically analyze the overall distribution of different viewing times, use an equal-frequency strategy to bin the viewing times, and determine the quantile of the distribution for each prediction sample, thereby predicting the viewing time to supplement the shortcomings of traditional video recommendation methods.

[0144] As can be seen from the figure, the model also uses data from external video streams (i.e., video streams in the first display mode) for training. The model consists of two parts. The input of the part below the dotted line in the figure is the data from the external video stream, which is used to train the video expression information of the trigger video in the external video stream. The trained video expression information is then input into the model above to achieve information exchange between the internal and external parts and provide the model with more knowledge. The portion above the dashed line in the diagram is used to estimate the estimated viewing time of candidate videos. The input to the network structure above is sample data from the internal video stream (i.e., the video stream in the second display mode). In feature construction, in addition to historical user interaction data between internal and external streams, video feature information of the candidate videos themselves, and user information of users in the internal stream, a new category of trigger video information is added: which external stream video generated the current internal stream consumption behavior. This allows the model to capture the implicit correlation between the estimated candidate videos and the trigger videos. Furthermore, an attention model structure is added based on user behavior and video information. After passing through a fully connected network layer, the output is a predicted value, estimating the viewing time distribution of candidate videos. This predicted value is limited to between 0 and 1. The model is trained using regression, and the regression loss (Huber loss) is calculated using the inverse gradient method. This model design enables data migration and correlation consideration in relevant business scenarios, making the model prediction more accurate and allowing for integration across different scenarios.

[0145] (2) Based on the matching of candidate videos and trigger videos in terms of identification information, classification information and tag information, each candidate video is divided into multiple preset empty queues to obtain multiple candidate video sequences.

[0146] (3) Within each candidate video sequence, sort each candidate video in order of the estimated viewing time length of each candidate video from long to short, to obtain a plurality of sorted candidate video sequences.

[0147] (4) Selecting recommended videos: first, select the candidate videos in the highest priority sorted candidate video sequence, and after all the candidate videos in the highest priority sorted candidate video sequence are selected, then select the candidate videos in the second highest priority sorted candidate video sequence, and finally select the candidate videos in the lowest priority sorted candidate video sequence. And when selecting in each sorted candidate video sequence, select in order of the estimated viewing time length from long to short.

[0148] (5) According to the order in which each recommended video is selected, form a recommended video sequence of each recommended video and recommend it to the user who selected the trigger video.

[0149] Through experiments, the method introduces video viewing time length to determine the recommended video, which increases the video playback time length of the detail page recommendation system by 1.2%. Through the improvement of the correlation between the trigger video and the candidate video, the detail page recommendation has been greatly improved in the associated action, and the similar score high proportion has increased by 0.3%. Using waterfall flow and detail page sliding form, from the long-term index, it can increase the use time length of the application program by 1%+, and the daily active user number (DAU) index also has a significant positive effect, which can show that through the form of internal and external linkage, it brings better consumption and novel browsing form and diverse interactive experience to the user, thereby improving the statistical data.

[0150] The video recommendation method provided by the embodiment considers the correlation with the current video in the recommendation system of the video detail page, and considers the correlation in each recommended funnel stage. First, in the recommendation recall stage, a separate trigger video related video trigger source is provided to ensure the underlying related supply. Secondly, in the precision ranking modeling stage, the trigger video features and the current video features are modeled to a certain extent, to ensure that the sorting model captures complete system knowledge. And in the final multi-objective sorting stage, the correlation is fully utilized for the sorting stage, so as to realize the internal and external double-flow recommendation linkage. On the one hand, the features of the external flow video (obtained by estimating the click rate) are added to the internal flow sorting model, and the internal flow estimated time length is added; on the other hand, in the scattering stage, the matching relationship between the type of the internal flow video and the trigger of the external flow click is judged for scattering, and the correlation consideration is appropriately weakened in the continuous sliding process of the user, to prevent the user from feeling fatigue, provide a different experience, and finally improve the browsing time length.

[0151] It should be understood that although each step in the flowchart involved in each embodiment as described above is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise specified herein, there is no strict order limitation for the execution of these steps, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be alternately or alternately executed with at least part of other steps or steps or stages in other steps.

[0152] It can be understood that the same / similar parts between each embodiment of the above method in the specification can be referred to each other, and each embodiment focuses on the difference from other embodiments, and the related part can be referred to the description of other method embodiments.

[0153] Based on the same inventive concept, the embodiments of the disclosure also provide a content recommendation device for implementing the above-mentioned content recommendation method, and a content recommendation model training device for implementing the above-mentioned content recommendation model training method.

[0154] Figure 5 is a structural block diagram of a content recommendation device according to an exemplary embodiment. Referring to Figure 5 The device includes an acquisition unit 510, a determination unit 520, a prediction unit 530, and a recommendation unit 540, wherein:

[0155] The acquisition unit 510 is configured to acquire candidate content associated with the trigger content and object information of the first type of object that has interacted with the trigger content; the first type of object is an object that performs an interaction behavior in a first display mode, and the first display mode is a content display mode before the trigger content is triggered;

[0156] The determination unit 520 is configured to obtain first content feature information of the candidate content in the first display mode and first object feature information of the first type of object in the first display mode by a first interaction prediction model corresponding to the first display mode; the first interaction prediction model is used to determine first interaction prediction information of the to-be-recommended content relative to the first type of object in the first display mode;

[0157] The prediction unit 530 is configured to determine second interaction prediction information of the to-be-recommended object to the candidate content in a second display mode based on content feature information of the trigger content, the first content feature information, and the first object feature information; the second display mode is a content display mode entered after the trigger content is triggered.

[0158] The recommendation unit 540 is configured to determine the recommended content from the candidate content based on the second interaction prediction information, and push the recommended content to the to-be-recommended object in the second display mode.

[0159] In an example embodiment, the prediction unit 530 is further configured to obtain at least one of historical interaction information of the to-be-recommended object for the content in the second display mode, content feature information of the candidate content, and object feature information of a second type of object that has interacted with the candidate content; the second type of object is an object that performs an interaction behavior in the second display mode; and determine the second interaction prediction information of the to-be-recommended object for the candidate content in the second display mode based on the at least one information, and the content feature information of the trigger content, the first content feature information, and the first object feature information.

[0160] In an example embodiment, the prediction unit 530 is further configured to, when the at least one information includes the historical interaction information and the content feature information of the candidate content, perform feature extraction processing on the historical interaction information and the content feature information of the candidate content by the attention unit to obtain extracted feature information; and obtain the second interaction prediction information of the to-be-recommended object for the candidate content based on the extracted feature information, the content feature information of the candidate content, and the content feature information of the trigger content, the first content feature information, and the object feature information of the first type of object.

[0161] In an example embodiment, the recommendation unit 540 is further configured to perform queue division on the candidate content according to matching information of the candidate content and the trigger content on a plurality of preset attribute information to obtain a plurality of candidate content sequences; and sort the candidate content in each candidate content sequence according to the second interaction prediction information to obtain a plurality of sorted candidate content sequences; and determine the recommended content from at least one of the sorted candidate content sequences according to a priority order of each candidate content sequence.

[0162] In an example embodiment, the recommendation unit 540 is further configured to obtain a plurality of empty queues; the empty queues are determined based on a priority order of each attribute information and a matching number on the attribute information; and divide each candidate content into each empty queue according to matching information corresponding to the candidate content and the trigger content and the priority order of each attribute information to obtain a plurality of candidate content sequences.

[0163] In an example embodiment, the recommendation unit 540 is further configured to obtain a target number of required recommended content; and select candidate content corresponding to the target number from at least one of the sorted candidate content sequences as the recommended content in a priority order of each candidate content sequence from high to low.

[0164] In an example embodiment, the prediction unit 530 is further configured to perform information prediction processing on the content feature information of the trigger content, the first content feature information, and the first object feature information by a second interaction prediction model corresponding to a second display mode, to obtain second interaction prediction information of the to-be-recommended object on the candidate content in the second display mode.

[0165] Figure 6 is a structural block diagram of a content recommendation device according to an example embodiment. Referring to Figure 6 The device includes an acquisition unit 610, a prediction unit 620, and a training unit 630, wherein:

[0166] The acquisition unit 610 is configured to perform acquisition of sample interaction data; the sample interaction data includes sample content feature information of sample trigger content, first sample content feature information of sample candidate content associated with the sample trigger content in a first display mode, second interaction information of a second type of sample object on the sample candidate content, and first sample object feature information of a first type of sample object that has interacted with the sample trigger content in the first display mode; the first sample content feature information and the first sample object feature information are obtained by a first interaction prediction model corresponding to the first display mode;

[0167] The prediction unit 620 is configured to perform information prediction processing on the sample content feature information, the first sample content feature information, and the first sample object feature information by a second interaction prediction model to be trained, to obtain second interaction prediction information of the second type of sample object on the sample candidate content in a second display mode;

[0168] The training unit 630 is configured to perform training of the second interaction prediction model to be trained based on a loss value between the second interaction prediction information and the second interaction information, to obtain a trained second interaction prediction model as a content recommendation model;

[0169] The first display mode is a content display mode before the sample trigger content triggers, the second display mode is a content display mode entered after the sample trigger content triggers, the first type of sample object is an object that performs an interaction behavior in the first display mode, the second type of sample object is an object that performs an interaction behavior in the second display mode, and the first interaction prediction model is used to determine first interaction prediction information of a to-be-recommended content relative to the first type of object in the first display mode.

[0170] Regarding the device in the above embodiments, the specific manner in which each module performs operations has been described in detail in the embodiments related to the method, and will not be described in detail here.

[0171] Each of the modules in the content recommendation apparatus described above can be implemented in whole or in part by software, hardware, and combinations thereof. The modules described above can be embedded in hardware in the form of a processor in the computer device or independent of the processor in the computer device, or can be stored in the form of software in the memory in the computer device so as to be called and executed by the processor.

[0172] Figure 7 FIG. 7 is a block diagram of an electronic device 700 for implementing a content recommendation method according to an exemplary embodiment. The electronic device 700 can be, for example, a mobile phone, a computer, a digital broadcasting terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, or the like.

[0173] Referring to Figure 7 The electronic device 700 can include one or more of the following components: a processing component 702, a memory 704, a power supply component 706, a multimedia component 708, an audio component 710, an input / output (I / O) interface 712, a sensor component 714, and a communication component 716.

[0174] The processing component 702 generally controls the overall operations of the electronic device 700, such as operations associated with display, phone calls, data communications, camera operations, and recording operations. The processing component 702 can include one or more processors 720 to execute instructions to complete all or part of the steps of the methods described above. In addition, the processing component 702 can include one or more modules to facilitate the interaction between the processing component 702 and other components. For example, the processing component 702 can include a multimedia module to facilitate the interaction between the multimedia component 708 and the processing component 702.

[0175] The memory 704 is configured to store various types of data to support operations of the electronic device 700. Examples of such data include instructions for any application or method operating on the electronic device 700, contact data, phonebook data, messages, pictures, videos, and the like. The memory 704 can be implemented by any type of volatile or non-volatile memory devices or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disc, optical disc, or graphene memory.

[0176] The power supply component 706 supplies power for the various components of the electronic device 700. The power supply component 706 can include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the electronic device 700.

[0177] The multimedia component 708 includes a screen to provide an output interface between the electronic device 700 and a user. In some embodiments, the screen can include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen can be implemented as a touch screen to receive an input signal from a user. The touch panel includes one or more touch sensors to sense a touch, a slide, and a gesture on the touch panel. The touch sensor can not only sense a boundary of a touching or a sliding action, but also detect duration and intensity of the touching or sliding action. In some embodiments, the multimedia component 708 includes a front camera and / or a rear camera. The front camera and / or the rear camera can receive external multimedia data when the electronic device 700 is in an operation mode, such as a shooting mode or a video mode. Each of the front and rear camera can be a fixed optical lens system or have a focal length and optical zoom ability.

[0178] The audio component 710 is configured to output and / or input an audio signal. For example, the audio component 710 includes a microphone (MIC) configured to receive an external audio signal when the electronic device 700 is in an operation mode, such as a call mode, a recording mode, and a voice recognition mode. The received audio signal can be further stored in the memory 704 or transmitted via the communication component 716. In some embodiments, the audio component 710 also includes a speaker to output an audio signal.

[0179] The I / O interface 712 provides an interface between the processing component 702 and peripheral interface modules, which can be a keypad, a click wheel, buttons, and the like. The buttons can include, but are not limited to, a home button, a volume button, a start button, and a lock button.

[0180] The sensor component 714 includes one or more sensors to provide various state assessments for the electronic device 700. For example, the sensor component 714 can detect an open / closed state of the electronic device 700, relative positioning of components, such as a display and a keypad of the electronic device 700, a change in position of the electronic device 700 or components of the electronic device 700, presence or absence of user contact with the electronic device 700, a change in orientation of the device 700 or acceleration / deceleration of the electronic device 700, and a temperature change of the electronic device 700. The sensor component 714 can include a proximity sensor configured to detect presence of a nearby object without any physical touch. The sensor component 714 can also include a light sensor such as a CMOS or CCD image sensor for use in an imaging application. In some embodiments, the sensor component 714 can further include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0181] The communication component 716 is configured to facilitate wired or wireless communication between the electronic device 700 and other devices. The electronic device 700 can access a wireless network based on a communication standard, such as WiFi, a cellular network standard (such as 2G, 3G, 4G, or 5G), or a combination thereof. In an exemplary embodiment, the communication component 716 receives broadcast signals or broadcast-related information from external broadcast management systems via a broadcast channel. In an exemplary embodiment, the communication component 716 further includes a Near Field Communication (NFC) module to facilitate close proximity communication. For example, the NFC module can be implemented based on Radio Frequency Identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0182] In an exemplary embodiment, the electronic device 700 can be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, micro-controllers, microprocessors, or other electronic elements, for performing the above-described methods.

[0183] In an exemplary embodiment, a computer-readable storage medium, such as the memory 704 including instructions, is also provided. The instructions can be executed by the processor 720 of the electronic device 700 to complete the above-described methods. For example, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disc, and an optical data storage device, etc.

[0184] In an exemplary embodiment, a computer program product including instructions is also provided. The instructions can be executed by the processor 720 of the electronic device 700 to complete the above-described methods.

[0185] It should be noted that the above-described apparatus, electronic device, computer-readable storage medium, computer program product, etc. according to the description of the method embodiments can also include other implementation manners, and the specific implementation manners can be referred to the description of the related method embodiments, which will not be repeated here.

[0186] Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the features of the disclosure disclosed herein. It is intended that the specification and examples be considered as exemplary only, with the true scope and spirit of the disclosure being indicated by the following claims.

[0187] It should be understood that the present disclosure is not limited to the precise construction that has been described above and shown in the accompanying drawings, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the present disclosure. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A content recommendation method characterized by, The method comprises: obtaining candidate content associated with trigger content and object information of a first type of object that has interacted with the trigger content; the first type of object is an object that performs an interaction behavior in a first display mode, and the first display mode is a content display mode before the trigger content is triggered; obtaining first content feature information of the candidate content in the first display mode and first object feature information of the first type of object in the first display mode through a first interaction prediction model corresponding to the first display mode; the first interaction prediction model is used to determine first interaction prediction information of a to-be-recommended content relative to the first type of object in the first display mode; determining second interaction prediction information of a to-be-recommended object to the candidate content in a second display mode based on content feature information of the trigger content, the first content feature information, and the first object feature information; the second display mode is a content display mode entered after the trigger content is triggered; determining a recommended content from the candidate content based on the second interaction prediction information, and pushing the recommended content to the to-be-recommended object in the second display mode.

2. The method of claim 1, wherein, The method comprises: obtaining at least one of historical interaction information of the to-be-recommended object to content in the second display mode, content feature information of the candidate content, and object feature information of a second type of object that has interacted with the candidate content; the second type of object is an object that performs an interaction behavior in the second display mode; determining second interaction prediction information of the to-be-recommended object to the candidate content in the second display mode based on the at least one information, and the content feature information of the trigger content, the first content feature information, and the first object feature information.

3. The method of claim 2, wherein, The method comprises: when the at least one information includes the historical interaction information and the content feature information of the candidate content, performing feature extraction processing on the historical interaction information and the content feature information of the candidate content through an attention unit to obtain extracted feature information; obtaining the second interaction prediction information of the to-be-recommended object to the candidate content based on the extracted feature information, the content feature information of the candidate content, and the content feature information of the trigger content, the first content feature information, and the object feature information of the first type of object.

4. The method of claim 1, wherein, The method comprises: According to the matching information of the candidate content and the trigger content on the preset attribute information, the candidate content is queue divided to obtain a plurality of candidate content sequences; According to the second interaction prediction information, the candidate content in each candidate content sequence is sorted to obtain a plurality of sorted candidate content sequences; According to the priority order of each candidate content sequence, the recommended content is determined from at least one sorted candidate content sequence.

5. The method of claim 4, wherein, According to the matching information of the candidate content and the trigger content on the preset attribute information, the candidate content is queue divided to obtain a plurality of candidate content sequences, including: Obtain a plurality of empty queues; the empty queue is determined based on the priority order of each attribute information and the matching number on the attribute information; According to the matching information of the candidate content and the trigger content, and the priority order of each attribute information, each candidate content is divided into each empty queue to obtain a plurality of candidate content sequences.

6. The method of claim 5, wherein, According to the priority order of each candidate content sequence, the recommended content is determined from at least one sorted candidate content sequence, including: Obtain the target number of required recommended content; According to the priority order of each candidate content sequence from high to low, select the candidate content corresponding to the target number from at least one sorted candidate content sequence as the recommended content.

7. The method of claim 1, wherein, Based on the content feature information of the trigger content, the first content feature information, and the first object feature information, the second interaction prediction information of the candidate content to be recommended object in the second display mode is determined, including: Through the second interaction prediction model corresponding to the second display mode, the content feature information of the trigger content, the first content feature information, and the first object feature information are information prediction processing, to obtain the second interaction prediction information of the candidate content to be recommended object in the second display mode. 8.A method for training a content recommendation model, the method comprising: Including: Obtain sample data; the sample data includes sample content feature information of sample trigger content, first sample content feature information of sample candidate content associated with the sample trigger content in the first display mode, second interaction information of the second sample object to the sample candidate content, and first sample object feature information of the first sample object in the first display mode; The first sample content feature information and the first sample object feature information are predicted by the first interaction prediction model corresponding to the first display mode; Through the second interaction prediction model to be trained, the sample content feature information, the first sample content feature information and the first sample object feature information are information prediction processing, to obtain the second interaction prediction information of the second sample object to the sample candidate content in the second display mode; train the second interaction prediction model to be trained based on a loss value between the second interaction prediction information and the second interaction information, to obtain a trained second interaction prediction model as the content recommendation model; The first display mode is a content display mode before the sample trigger content is triggered, the second display mode is a content display mode entered after the sample trigger content is triggered, the first type of sample object is an object performing an interaction behavior in the first display mode, the second type of sample object is an object performing an interaction behavior in the second display mode, and the first interaction prediction model is used to determine first interaction prediction information of the to-be-recommended content relative to the first type of sample object in the first display mode.

9. A content recommendation apparatus characterized by comprising: Comprise: The acquisition unit is configured to acquire candidate content associated with trigger content and object information of a first type of object that has interacted with the trigger content; The first type of object is an object performing an interaction behavior in a first display mode, and the first display mode is a content display mode before the trigger content is triggered; The determination unit is configured to obtain first content feature information of the candidate content in the first display mode and first object feature information of the first type of object in the first display mode by using a first interaction prediction model corresponding to the first display mode; The first interaction prediction model is used to determine first interaction prediction information of to-be-recommended content relative to the first type of object in the first display mode; The prediction unit is configured to determine second interaction prediction information of a to-be-recommended object to the candidate content in a second display mode based on content feature information of the trigger content, the first content feature information, and the first object feature information; and the second display mode is a content display mode entered after the trigger content is triggered. The recommendation unit is configured to determine recommended content from the candidate content based on the second interaction prediction information, and push the recommended content to the to-be-recommended object in the second display mode.

10. The apparatus of claim 9, wherein, The prediction unit is further configured to acquire at least one of historical interaction information of the to-be-recommended object to content in the second display mode, content feature information of the candidate content, and object feature information of a second type of object that has interacted with the candidate content; The second type of object is an object performing an interaction behavior in the second display mode; and the second interaction prediction information of the to-be-recommended object to the candidate content in the second display mode is determined based on the at least one information, and the content feature information of the trigger content, the first content feature information, and the first object feature information.

11. The apparatus of claim 10, wherein, The prediction unit is further configured to perform, when the history interaction information and the content feature information of the candidate content are included in the at least one information, feature extraction processing on the history interaction information and the content feature information of the candidate content by an attention unit to obtain extracted feature information; and obtain second interaction prediction information of the to-be-recommended object on the candidate content based on the extracted feature information, the content feature information of the candidate content, and the content feature information of the trigger content, the first content feature information, and the object feature information of the first type of object.

12. The apparatus of claim 9, wherein, The recommendation unit is further configured to perform queue division on the candidate content according to the matching information of the candidate content and the trigger content on the preset multiple attribute information to obtain multiple candidate content sequences. According to the second interaction prediction information, the candidate contents in each candidate content sequence are sorted respectively to obtain multiple sorted candidate content sequences; and the recommendation content is determined from at least one of the sorted candidate content sequences according to the priority order of each candidate content sequence.

13. The apparatus of claim 12, wherein, The recommendation unit is further configured to perform obtaining a plurality of empty queues; the empty queue is determined based on the priority order of each attribute information and the matching number on the attribute information; and each candidate content is divided into each empty queue according to the matching information corresponding to the candidate content and the trigger content and the priority order of each attribute information to obtain multiple candidate content sequences.

14. The apparatus of claim 13, wherein, The recommendation unit is further configured to perform obtaining a target number of required recommendation content; and selecting candidate content corresponding to the target number from at least one of the sorted candidate content sequences as recommendation content in the order from high to low of the priority of each candidate content sequence.

15. The apparatus of claim 9, wherein, The prediction unit is further configured to perform information prediction processing on the content feature information of the trigger content, the first content feature information, and the first object feature information by a second interaction prediction model corresponding to the second display mode to obtain second interaction prediction information of the to-be-recommended object on the candidate content in the second display mode.

16. A training device for a content recommendation model, characterized in that, The obtaining unit is configured to perform obtaining sample interaction data; the sample interaction data includes sample content feature information of a sample trigger content, first sample content feature information of a sample candidate content associated with the sample trigger content in a first display mode, second interaction information of a second type of sample object on the sample candidate content, and first sample object feature information of a first type of sample object that has interacted with the sample trigger content in the first display mode; The first sample content feature information and the first sample object feature information are predicted by a first interaction prediction model corresponding to the first display mode; ​ a prediction unit configured to perform information prediction processing on the sample content feature information, the first sample content feature information, and the first sample object feature information by a second interaction prediction model to be trained, to obtain second interaction prediction information of a second type of sample object on the sample candidate content in a second display mode; a training unit configured to train the second interaction prediction model to be trained based on a loss value between the second interaction prediction information and the second interaction information, to obtain a trained second interaction prediction model as a content recommendation model; wherein the first display mode is a content display mode before the sample trigger content is triggered, the second display mode is a content display mode entered after the sample trigger content is triggered, the first type of sample object is an object performing an interaction behavior in the first display mode, the second type of sample object is an object performing an interaction behavior in the second display mode, and the first interaction prediction model is used to determine first interaction prediction information of a to-be-recommended content relative to the first type of sample object in the first display mode.

17. An electronic device, comprising: comprise: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the content recommendation method of any one of claims 1 to 7 or the training method of the content recommendation model of claim 8.

18. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is enabled to perform the content recommendation method of any one of claims 1 to 7 or the training method of the content recommendation model of claim 8.

19. A computer program product, comprising instructions therein, characterised in that, When the instructions are executed by the processor of the electronic device, the electronic device is enabled to perform the content recommendation method of any one of claims 1 to 7 or the training method of the content recommendation model of claim 8.

Images