CN114154068A

CN114154068A - Media content recommendation method and device, electronic equipment and storage medium

Info

Publication number: CN114154068A
Application number: CN202111481105.XA
Authority: CN
Inventors: 王朝坤; 吴呈; 徐劲草; 王昶平; 宋洋
Original assignee: Tsinghua University; Beijing Dajia Internet Information Technology Co Ltd
Current assignee: Tsinghua University; Beijing Dajia Internet Information Technology Co Ltd
Priority date: 2021-12-06
Filing date: 2021-12-06
Publication date: 2022-03-08

Abstract

The disclosure relates to a media content recommendation method, a media content recommendation device, electronic equipment and a storage medium, and belongs to the technical field of computers. The method comprises the following steps: acquiring a recommendation model, wherein the recommendation model comprises object characteristics of a plurality of objects, and the plurality of objects comprise user accounts and media contents; determining similarity between the first user account and the first media content based on account characteristics of the first user account and content characteristics of the first media content; determining to recommend the first media content to the first user account based on the similarity; the plurality of objects includes a first object and a second object, and the recommendation model is trained based on propagation loss inversely correlated to a first similarity determined based on a similarity between the influencing features of the second object and the object features of the second object. The object features in the recommendation model obtained by training in the method are accurate, so that the similarity determined based on the recommendation model is accurate, and the recommendation accuracy of the media content is improved.

Description

Media content recommendation method and device, electronic equipment and storage medium

Technical Field

The present disclosure relates to the field of computer technologies, and in particular, to a method and an apparatus for recommending media content, an electronic device, and a storage medium.

Background

With the development of the internet and the wide application of recommendation technologies, recommending media content to an account based on a recommendation model has become a common recommendation method. The recommendation model will typically make recommendations based on characteristics of the account number and characteristics of the media content. Wherein the features are obtained by training the recommendation model based on the behavior information. The behavior information includes at least one piece of behavior data, and each piece of behavior data can represent an interaction behavior between one piece of media content and one account, such as a behavior of displaying the media content based on the account, or a behavior of forwarding the media content based on the account, and the like. The recommendation model is capable of determining characteristics of each media content and each account based on the at least one piece of behavior data.

However, how to improve the recommendation accuracy of the recommendation model based on the new interaction behavior between the account and the media content becomes an urgent problem to be solved.

Disclosure of Invention

The disclosure provides a media content recommendation method, a media content recommendation device, an electronic device and a storage medium, which improve recommendation accuracy.

According to an aspect of an embodiment of the present disclosure, there is provided a media content recommendation method, the method including:

acquiring a recommendation model, wherein the recommendation model comprises object characteristics of a plurality of objects, and the plurality of objects comprise user accounts and media contents;

determining similarity between a first user account and first media content based on account characteristics of the first user account and content characteristics of the first media content, wherein the first user account is any one of the user accounts, and the first media content is any one of the media content;

determining to recommend the first media content to the first user account based on the similarity;

the plurality of objects comprise a first object and a second object, the first object is an object in the newly added behavior data, the second object is an object related to the first object, the recommendation model is obtained based on propagation loss training negatively related to first similarity, the first similarity is determined based on similarity between an influence feature of the second object and an object feature of the second object, and the influence feature represents a feature of the second object under the influence of the interaction behavior corresponding to the newly added behavior data.

In some embodiments, the account characteristics include first and second remembered characteristics of the first user account, the first remembered characteristic characterizing long-term characteristics of the first user account, the second remembered characteristic characterizing short-term characteristics of the first user account; the content features comprise first memory features and second memory features of the first media content, the first memory features of the first media content characterize the first media content in a long term, the second memory features of the first media content characterize the first media content in a short term;

the determining the similarity between the first user account and the first media content based on account characteristics of the first user account and content characteristics of the first media content includes:

fusing the first memory characteristic and the second memory characteristic of the first user account to obtain a fused characteristic of the first user account, and fusing the first memory characteristic and the second memory characteristic of the first media content to obtain a fused characteristic of the first media content;

determining a similarity between the fusion characteristics of the first user account and the fusion characteristics of the first media content.

In some embodiments, the account characteristics further include contextual characteristics of the first user account, the contextual characteristics characterizing characteristics of the first user account under the influence of other objects;

the fusing the first memory characteristics and the second memory characteristics of the first user account to obtain the fused characteristics of the first user account includes:

and fusing the first memory characteristic, the second memory characteristic and the context characteristic of the first user account to obtain a fused characteristic of the first user account.

According to still another aspect of the embodiments of the present disclosure, there is provided a recommendation model processing method, including:

acquiring first behavior information, wherein the first behavior information comprises newly added behavior data and historical behavior data, the newly added behavior data is used for representing interaction behavior between two first objects which belong to different types, the types comprise account numbers and media contents, and the historical behavior data is historical behavior data corresponding to a second object related to any one of the first objects;

obtaining a recommendation model comprising first features of a plurality of objects, the plurality of objects comprising the two first objects and at least one second object in the historical behavior data;

for each first object, determining the influence characteristics of the second object based on the first characteristics of the first object and the interaction time difference of the second object related to the first object, wherein the influence characteristics of any second object represent the characteristics of the second object under the influence of the interaction behavior corresponding to the newly added behavior data, and the interaction time difference of the second object is the time difference between the occurrence time of the newly added behavior data and the occurrence time of the historical behavior data to which the second object belongs;

training the recommendation model based on a propagation loss negatively correlated with a first similarity determined based on a similarity between the impact feature of each of the second objects and the first feature of each of the second objects, the trained recommendation model including the second features of the plurality of objects, and the trained recommendation model for recommending based on the second features of the plurality of objects.

In some embodiments, the determining, for each of the first objects, an impact feature of the second object based on a difference in interaction time of the first feature of the first object and the second object associated with the first object includes:

for each first object, attenuating the first feature of the first object based on the interaction time difference of the first object to obtain the interaction feature of the first object, wherein the interaction time difference of the first object is the time difference between the occurrence time of the newly added behavior data and the occurrence time of the historical behavior data to which the first object belongs;

determining an influence characteristic of the second object related to the first object based on an interaction characteristic of the first object and an interaction time difference of the second object belonging to the same historical behavior data as the first object;

and continuing to determine the influence characteristic of another second object based on the influence characteristic of the second object and the interaction time difference of the second object belonging to the same historical behavior data with the second object until the influence characteristic of each second object related to the first object in the first behavior information is determined.

In some embodiments, the determining, based on the interaction characteristics of the first object and the interaction time difference of the second object belonging to the same historical behavior data as the first object, the influence characteristics of the second object related to the first object includes:

under the condition that the interaction time difference is not larger than a time difference threshold value, determining a first attenuation parameter negatively correlated with the interaction time difference, and attenuating the interaction characteristic of the first object according to the first attenuation parameter to obtain an influence characteristic of the second object; alternatively, the first and second electrodes may be,

and determining a preset influence characteristic as the influence characteristic of the second object when the interaction time difference is larger than the time difference threshold value.

In some embodiments, the first behavior information includes at least two object nodes belonging to different node types and edges connecting between any two object nodes, and the node types include an account type and a media content type; wherein the at least two object nodes include two first object nodes belonging to different types and a second object node directly or indirectly connected to any one of the first object nodes;

the two first object nodes and a first edge connected between the two first object nodes form the newly added behavior data;

the first object node and the second object node which belong to different types and a second edge connected between the first object node and the second object node form historical behavior data, and/or any two second object nodes which belong to different types and a third edge connected between any two second object nodes form historical behavior data;

for each of the first objects, determining an influence characteristic of the second object based on a difference in interaction time of the first feature of the first object and the second object associated with the first object, including:

for each first object node, attenuating the first feature of the first object node based on the interaction time difference of the first object node to obtain the interaction feature of the first object node, wherein the interaction time difference of the first object node is the time difference between the occurrence time of the first edge and the occurrence time of a second edge connected with the first object node;

determining an influence characteristic of the second object node based on an interaction characteristic of the first object node and an interaction time difference of the second object node directly connected with the first object node;

and continuing to determine the influence characteristic of another second object node based on the influence characteristic of the second object node and the interaction time difference of another second object node directly connected with the second object node until determining the influence characteristic of each second object node directly or indirectly connected with any first object node in the first behavior information.

In some embodiments, the method further comprises:

for each first object, attenuating first features of the first object based on an interaction time difference of the first object to obtain attenuation features, wherein the attenuation features represent features of the first object after the first features of the first object are attenuated under the influence of an interaction behavior corresponding to the newly added behavior data, and the interaction time difference of the first object is a time difference between the occurrence time of the newly added behavior data and the occurrence time of historical behavior data to which the first object belongs;

determining an interaction loss negatively correlated with a second similarity based on the attenuation features of the two first objects, the second similarity being a similarity between the attenuation features of the two first objects;

training the recommendation model based on the propagation loss negatively correlated with the first similarity, including:

training the recommendation model based on the propagation loss and the interaction loss that are negatively correlated with the first similarity.

In some embodiments, the first feature of the first subject comprises a first memory feature characterizing long-term features of the first subject and a second memory feature characterizing short-term features of the first subject; the attenuating the first feature of the first object based on the interaction time difference of the first object to obtain an attenuated feature includes:

attenuating a second memory characteristic of the first object based on the interaction time difference of the first object;

and fusing the first memory characteristics of the first object and the second memory characteristics after attenuation to obtain the attenuation characteristics of the first object.

In some embodiments, said attenuating the second memory characteristic of the first object based on the interaction time difference of the first object comprises:

determining a second attenuation parameter based on the interaction time difference of the first object and a learning parameter corresponding to the type of the first object;

attenuating the second memory characteristic based on the second attenuation parameter.

In some embodiments, the first feature of the first object further comprises a contextual feature characterizing features of the first object under the influence of other objects; the fusing the first memory feature of the first object with the attenuated second memory feature to obtain the attenuated feature of the first object includes:

and performing weighted fusion on the first memory characteristic of the first object, the attenuated second memory characteristic and the context characteristic to obtain the attenuation characteristic of the first object.

In some embodiments, the context features of the first object include context features of the first object for a plurality of interaction types, and the new added behavior data includes a target interaction type corresponding to the interaction behavior;

the performing weighted fusion on the first memory feature, the attenuated second memory feature and the context feature of the first object to obtain the attenuated feature of the first object includes:

determining context characteristics corresponding to the target interaction type from the context characteristics of the first object aiming at multiple interaction types;

and performing weighted fusion on the first memory characteristic of the first object, the attenuated second memory characteristic and the context characteristic corresponding to the target interaction type to obtain the attenuation characteristic of the first object.

In some embodiments, the recommendation model further comprises model parameters, and the training of the recommendation model based on the propagation loss and the interaction loss negatively correlated with the first similarity comprises:

fusing the propagation loss and the interaction loss which are negatively correlated with the first similarity to obtain the model loss of the recommendation model;

and updating the model parameters in the recommendation model and the first characteristics of the plurality of objects based on the model loss to obtain the trained recommendation model.

In some embodiments, the method further comprises:

determining a third object corresponding to each first object from second behavior information, wherein the second behavior information comprises the newly added behavior data and a plurality of pieces of historical behavior data, and the third object is other objects except the first object in the second behavior information;

determining a negative sampling loss negatively correlated with a third similarity determined based on a similarity between the interaction feature of the first object and a context feature of the third object, the context feature characterizing features of the third object under influence of other objects;

training the recommendation model based on the propagation loss negatively correlated with the first similarity and the negative sampling loss.

In some embodiments, the method further comprises:

training the recommendation model based on the propagation loss and the interaction loss negatively correlated with the first similarity, including:

training the recommendation model based on the propagation loss, the interaction loss, and the negative sampling loss that are negatively correlated with the first similarity.

In some embodiments, the obtaining the first behavior information includes:

and sampling second behavior information according to the two first objects in the newly added behavior data to obtain the first behavior information, wherein the second behavior information comprises the newly added behavior data and a plurality of pieces of historical behavior data.

In some embodiments, the sampling second behavior information according to the two first objects in the newly added behavior data to obtain the first behavior information includes:

acquiring a sampling mode set, wherein the sampling mode set comprises a plurality of sampling modes;

determining a sampling mode of each first object from the sampling mode set;

for each first object, sampling is carried out in the second behavior information according to the determined sampling mode from the first object, and historical behavior data obtained by sampling and the newly added behavior data form the first behavior information.

In some embodiments, the first feature of the second object comprises a contextual feature characterizing features of the second object under the influence of other objects; the method further comprises the following steps:

determining a similarity between the impact feature and the context feature of each of the second objects;

and fusing the similarity of the at least one second object to obtain the first similarity.

In some embodiments, before the obtaining of the new behavior data, the method further includes:

obtaining a plurality of pieces of sample behavior data, wherein each piece of sample behavior data is used for representing interaction behavior between two sample objects which belong to different types;

dividing the plurality of pieces of sample behavior data into a plurality of sample sets according to the sequence of the occurrence time from far to near, wherein the number of the sample behavior data contained in each sample set is the same;

training the recommendation model in sequence based on the plurality of sample sets.

According to still another aspect of the embodiments of the present disclosure, there is provided a media content recommendation apparatus, the apparatus including:

a first obtaining unit configured to perform obtaining a recommendation model, the recommendation model including object characteristics of a plurality of objects, the plurality of objects including a user account and media content;

a first determining unit, configured to perform determining a similarity between a first user account and first media content based on account characteristics of the first user account and content characteristics of the first media content, where the first user account is any one of the user accounts, and the first media content is any one of the media content;

a recommending unit configured to perform recommending the first media content to the first user account based on the similarity determination;

the first determining unit is configured to perform fusion of the first memory feature and the second memory feature of the first user account to obtain a fusion feature of the first user account, and perform fusion of the first memory feature and the second memory feature of the first media content to obtain a fusion feature of the first media content; determining a similarity between the fusion characteristics of the first user account and the fusion characteristics of the first media content.

the first determining unit is configured to perform fusion of the first memory feature, the second memory feature and the context feature of the first user account to obtain a fusion feature of the first user account.

According to still another aspect of the embodiments of the present disclosure, there is provided a recommendation model processing apparatus, the apparatus including:

a second obtaining unit, configured to perform obtaining of first behavior information, where the first behavior information includes new behavior data and historical behavior data, the new behavior data is used to represent an interaction behavior between two first objects belonging to different types, the types include an account and media content, and the historical behavior data is historical behavior data corresponding to a second object related to any one of the first objects;

a third obtaining unit configured to perform obtaining a recommendation model, the recommendation model including first features of a plurality of objects, the plurality of objects including the two first objects and at least one second object in the historical behavior data;

a second determining unit, configured to perform, for each first object, determining influence characteristics of the second object based on interaction time differences between first characteristics of the first object and second objects related to the first object, where the influence characteristics of any second object characterize characteristics of the second object under influence of an interaction behavior corresponding to the new added behavior data, and the interaction time difference of the second object is a time difference between an occurrence time of the new added behavior data and an occurrence time of historical behavior data to which the second object belongs;

a training unit configured to perform training of the recommendation model based on a propagation loss negatively correlated with a first similarity determined based on a similarity between an influence feature of each of the second objects and a first feature of each of the second objects, the trained recommendation model including second features of the plurality of objects, and the trained recommendation model for recommendation based on the second features of the plurality of objects.

In some embodiments, the second determining unit is configured to perform, for each first object, attenuation on a first feature of the first object based on an interaction time difference of the first object, so as to obtain an interaction feature of the first object, where the interaction time difference of the first object is a time difference between an occurrence time of the new added behavior data and an occurrence time of historical behavior data to which the first object belongs;

In some embodiments, the second determining unit is configured to perform, in a case that the interaction time difference is not greater than a time difference threshold, determining a first attenuation parameter negatively correlated to the interaction time difference, and attenuating the interaction feature of the first object according to the first attenuation parameter to obtain the influence feature of the second object; alternatively, the first and second electrodes may be,

the second determination unit is configured to perform determining a preset influence characteristic as the influence characteristic of the second object if the interaction time difference is greater than the time difference threshold.

the second determining unit is configured to perform attenuation on the first feature of the first object node based on the interaction time difference of the first object node, so as to obtain the interaction feature of the first object node, where the interaction time difference of the first object node is a time difference between the occurrence time of the first edge and the occurrence time of a second edge connected to the first object node;

In some embodiments, the apparatus further comprises:

a third determining unit, configured to perform attenuation on first features of the first objects based on an interaction time difference of the first objects for each first object to obtain an attenuation feature, where the attenuation feature represents a feature of the first object after the first features of the first objects are attenuated under the influence of an interaction behavior corresponding to the newly added behavior data, and the interaction time difference of the first objects is a time difference between an occurrence time of the newly added behavior data and an occurrence time of historical behavior data to which the first objects belong;

the training unit is configured to perform training on the recommendation model based on the propagation loss and the interaction loss which are inversely related to the first similarity.

In some embodiments, the first feature of the first subject comprises a first memory feature characterizing long-term features of the first subject and a second memory feature characterizing short-term features of the first subject; the third determining unit is configured to perform attenuation of a second memory characteristic of the first object based on the interaction time difference of the first object; and fusing the first memory characteristics of the first object and the second memory characteristics after attenuation to obtain the attenuation characteristics of the first object.

In some embodiments, the third determining unit is configured to perform determining a second attenuation parameter based on the interaction time difference of the first object and a learning parameter corresponding to the type of the first object; attenuating the second memory characteristic based on the second attenuation parameter.

In some embodiments, the first feature of the first object further comprises a contextual feature characterizing features of the first object under the influence of other objects; the third determining unit is configured to perform weighted fusion on the first memory feature of the first object, the attenuated second memory feature and the context feature to obtain an attenuation feature of the first object.

the third determining unit is configured to determine, from the context features of the first object for multiple interaction types, a context feature corresponding to the target interaction type; and performing weighted fusion on the first memory characteristic of the first object, the attenuated second memory characteristic and the context characteristic corresponding to the target interaction type to obtain the attenuation characteristic of the first object.

In some embodiments, the recommendation model further includes model parameters, and the training unit is configured to perform fusion of the propagation loss and the interaction loss that are negatively correlated with the first similarity, so as to obtain a model loss of the recommendation model; and updating the model parameters in the recommendation model and the first characteristics of the plurality of objects based on the model loss to obtain the trained recommendation model.

In some embodiments, the apparatus further comprises:

a fourth determining unit configured to perform determining a third object corresponding to each of the first objects from second behavior information, the second behavior information including the newly added behavior data and a plurality of pieces of historical behavior data, the third object being an object other than the first object in the second behavior information; determining a negative sampling loss negatively correlated with a third similarity determined based on a similarity between the interaction feature of the first object and a context feature of the third object, the context feature characterizing features of the third object under influence of other objects;

the training unit is configured to perform training on the recommendation model based on the propagation loss negatively correlated with the first similarity and the negative sampling loss.

In some embodiments, the apparatus further comprises:

a fourth determining unit configured to perform determining a third object corresponding to each of the first objects from second behavior information, the second behavior information including the newly added behavior data and a plurality of pieces of historical behavior data, the third object being an object other than the first object in the second behavior information;

the training unit is configured to perform training on the recommendation model based on the propagation loss, the interaction loss and the negative sampling loss which are negatively correlated with the first similarity.

In some embodiments, the second obtaining unit is configured to perform sampling on second behavior information according to the two first objects in the new added behavior data to obtain the first behavior information, where the second behavior information includes the new added behavior data and a plurality of pieces of historical behavior data.

In some embodiments, the second obtaining unit is configured to perform obtaining a sampling mode set, where the sampling mode set includes a plurality of sampling modes; determining a sampling mode of each first object from the sampling mode set; for each first object, sampling is carried out in the second behavior information according to the determined sampling mode from the first object, and historical behavior data obtained by sampling and the newly added behavior data form the first behavior information.

In some embodiments, the first feature of the second object comprises a contextual feature characterizing features of the second object under the influence of other objects; the device further comprises:

a fifth determining unit configured to perform determining a similarity between the influence feature and the context feature of each of the second objects; and fusing the similarity of the at least one second object to obtain the first similarity.

In some embodiments, the apparatus further comprises:

the training unit is further configured to perform obtaining a plurality of pieces of sample behavior data, each piece of sample behavior data being used for representing an interaction behavior between two sample objects belonging to different types; dividing the plurality of pieces of sample behavior data into a plurality of sample sets according to the sequence of the occurrence time from far to near, wherein the number of the sample behavior data contained in each sample set is the same; training the recommendation model in sequence based on the plurality of sample sets.

According to still another aspect of the embodiments of the present disclosure, there is provided an electronic apparatus including:

one or more processors;

a memory for storing the one or more processor-executable instructions;

wherein the one or more processors are configured to perform the media content recommendation method or recommendation model processing method of the above aspects.

According to still another aspect of the embodiments of the present disclosure, there is provided a computer-readable storage medium, wherein instructions of the storage medium, when executed by a processor of an electronic device, enable the electronic device to perform the media content recommendation method or the recommendation model processing method of the above aspect.

According to still another aspect of the embodiments of the present disclosure, there is provided a computer program product comprising a computer program executed by a processor to implement the media content recommendation method or the recommendation model processing method of the above aspect.

In the embodiment of the disclosure, the newly added behavior data represents a new interactive behavior generated by the first object, the interactive behavior may affect a second object related to the first object, and the influence characteristics of the second object affected by the interactive behavior corresponding to the newly added behavior data are determined, so that training is performed based on the influence characteristics, and thus, the accuracy of object characteristics included in the trained recommendation model is higher, the accuracy of similarity between the first user account and the first media content determined based on the recommendation model is higher, and the recommendation accuracy of the media content is further improved.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and together with the description, serve to explain the principles of the disclosure.

FIG. 1 is a schematic diagram illustrating one implementation environment in accordance with an example embodiment.

FIG. 2 is a flow diagram illustrating a recommendation model processing method in accordance with an exemplary embodiment.

FIG. 3 is a flow diagram illustrating a recommendation model processing method in accordance with an exemplary embodiment.

FIG. 4 is a diagram illustrating a training process for a recommendation model, according to an example embodiment.

FIG. 5 is a flow diagram illustrating a recommendation model processing method in accordance with an exemplary embodiment.

FIG. 6 is a flow chart illustrating a method of media content recommendation, according to an example embodiment.

FIG. 7 is a flow chart illustrating a method of media content recommendation, according to an example embodiment.

FIG. 8 is a block diagram illustrating a media content recommender, according to an exemplary embodiment.

Fig. 9 is a block diagram illustrating a recommendation model processing apparatus according to an example embodiment.

Fig. 10 is a block diagram illustrating a structure of a terminal according to an exemplary embodiment.

FIG. 11 is a block diagram illustrating the structure of a server in accordance with an exemplary embodiment.

Detailed Description

In order to make the technical solutions of the present disclosure better understood by those of ordinary skill in the art, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

It should be noted that the terms "first," "second," and the like in the description and claims of the present disclosure and in the description of the above-described figures are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the disclosure described herein are capable of operation in sequences other than those illustrated or otherwise described herein. The implementations described in the exemplary embodiments below are not intended to represent all implementations consistent with the present disclosure. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present disclosure, as detailed in the appended claims.

It should be noted that, as used in this disclosure, the terms "at least one," "a plurality," "each," "any," and the like, at least one includes one, two, or more than two, and a plurality includes two or more than two, each referring to each of the corresponding plurality, and any referring to any one of the plurality.

It should be noted that the user data (including but not limited to user device information, user personal information, etc.) referred to in the present disclosure is information authorized by the user or sufficiently authorized by each party.

The execution subject of the media content recommendation method provided by the embodiment of the disclosure is electronic equipment. Optionally, the electronic device is a terminal or a server, and the media content recommendation method can be implemented by the terminal or the server, or by interaction between the terminal and the server, which is not limited in this disclosure. In the embodiment of the present disclosure, a method for implementing media content recommendation through interaction between a terminal and a server is described as an example.

FIG. 1 is a schematic diagram of an implementation environment, shown in accordance with an exemplary embodiment, and referring to FIG. 1, the implementation environment includes: a terminal 110 and a server 120. The terminal 110 is connected to the server 120 through a wireless network or a wired network.

Optionally, the terminal 110 is a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto. The terminal 110 may be generally referred to as one of a plurality of terminals, and the embodiment is only illustrated by the terminal 110. Those skilled in the art will appreciate that the number of terminals described above may be greater or fewer. In some embodiments, the terminal 110 is installed with a media content presentation application that is served by the server 120. The terminal 110 can implement data interaction with the server 120 through the media content presentation application. The media content presentation application may be a video application or a music application, etc.

Optionally, the server 120 is a server, a server cluster composed of several servers, or a cloud computing service center. The number of servers 120 may be more or less, and the embodiment of the disclosure does not limit this. Of course, the server 120 may also include other functional servers in order to provide more comprehensive and diversified services.

In the embodiment of the present disclosure, a user performs one or more behaviors on media content on a terminal 110, the terminal 110 logs in an account, so that an interactive behavior is generated between the account and the media content, the terminal 110 acquires behavior data corresponding to the interactive behavior, sends the behavior data to a server 120, and the server 102 trains a recommendation model based on the behavior data. The server 120 determines the media content recommended to the account based on the trained recommendation model, the server 120 sends the media content to the terminal 110 logged in to the account, and the terminal 110 displays the media content so that the user operating the terminal 110 can view the media content.

It should be noted that, in the embodiment of the present disclosure, the behavior data used for training the recommendation model can be uploaded to the server by the terminal, and can also be obtained by the server itself, which is not limited in the embodiment of the present disclosure.

After the implementation environment of the embodiment of the present disclosure is described, an application scenario of the embodiment of the present disclosure will be described below with reference to the implementation environment. It should be noted that, in the following description, a terminal is also the terminal 110, and a server is also the server 120.

In some embodiments, the method provided by the embodiments of the present disclosure can be applied in a recommendation scene of media content. The media content includes video, pictures or audio, etc. Taking media content as a short video as an example, when a user browses the short video through a terminal, interaction behaviors between an account and the short video are recorded as new behavior data, such as praise behaviors, comment behaviors or forwarding behaviors. The server trains the recommendation model based on the newly added behavior data by adopting the recommendation model processing method provided by the embodiment of the disclosure. When subsequently recommending the account logged in the terminal, the server determines the short video recommended for the account according to the trained recommendation model by adopting the media content recommendation method provided by the embodiment of the disclosure, so as to recommend the short video for the account.

In other embodiments, the method provided by the embodiment of the present disclosure can also be applied to other recommendation scenarios, for example, an item recommendation scenario, where when a user transacts an item through a terminal, an interaction behavior between an account and the item is recorded as new-added behavior data, such as a transaction behavior, a viewing behavior, or a collection behavior. The server trains the recommendation model by adopting the recommendation model processing method provided by the embodiment of the disclosure based on the newly added behavior data. When subsequently recommending the account logged in the terminal, the server determines the item recommended for the account according to the trained recommendation model by using the media content recommendation method provided by the embodiment of the disclosure, so as to recommend the item for the account.

It should be noted that the recommendation model processing method provided in the embodiment of the present disclosure can also be applied to other scenarios of processing a model, and the embodiment of the present disclosure does not limit this scenario.

Fig. 2 is a flowchart illustrating a recommendation model processing method according to an exemplary embodiment, and referring to fig. 2, the method is executed by an electronic device, and includes the following steps:

in step 201, the electronic device obtains first behavior information, where the first behavior information includes new behavior data and historical behavior data, the new behavior data is used to represent an interaction behavior between two first objects that belong to different types, the types include an account and media content, and the historical behavior data is historical behavior data corresponding to a second object related to any one of the first objects.

In some embodiments, one piece of added behavior data is data corresponding to an interaction behavior of the account on the media content. The historical behavior data is data corresponding to the interaction behavior of the account on the media content before the new behavior data is generated. The historical behavior data included in the first behavior information may be regarded as historical behavior data that is affected by the newly added behavior data.

Optionally, the interactive behavior includes a praise behavior, a forward behavior, a publish behavior, a collect behavior, or a comment behavior. Correspondingly, the type to which the interactive behavior belongs, that is, the interaction type, includes a plurality of types, and the interaction type includes a like type, a forwarding type, a publishing type, a collection type, a comment type, or the like.

In some embodiments, one account has at least one account type. That is, the account type of the account is at least one of a user type and an author type. The author account is an account for publishing media content, and the user account is an account for interacting with media content published by other accounts. In this embodiment of the present disclosure, since a piece of new behavior data is generated for an interactive behavior, the account type of the account performing the interactive behavior is unique, for example, if the account issues media content, the account is an author account, and if the account approves the media content, the account is a user account. That is, the account performing the interactive behavior in the newly added behavior data may be a user account or an author account.

In step 202, the electronic device obtains a recommendation model, the recommendation model including first features of a plurality of objects, the plurality of objects including two first objects and at least one second object in historical behavior data.

Because one account may have two account types, that is, two attributes, and each piece of newly added behavior data is for an account of one account type, for accounts of two account types, the recommendation model may process for an author account and for a user account, respectively, so that the trained recommendation model includes two features of the account, which are features of the author account and features of the user account, respectively, thereby distinguishing features of the same account in different account types. The characteristics of the author account are the characteristics of the account under the author attribute, and the characteristics of the user account are the characteristics of the account under the user attribute.

The first feature of the object can represent semantic information of the media content matched by the object, the semantic information being related information of the media content learned in the training process of the recommendation model, e.g., the semantic information can represent a type of the media content or other information. The first characteristic of the account represents semantic information of the media content matched with the account, and the first characteristic of the media content represents the semantic information of the media content. The matched media content refers to the media content preferred by the account.

In step 203, the electronic device determines, for each first object, an impact feature of a second object based on a difference in interaction time of the first feature of the first object and the second object associated with the first object.

The influence characteristics of any second object represent the characteristics of the second object under the influence of the interaction behavior corresponding to the newly added behavior data, and the interaction time difference of the second object is the time difference between the occurrence time of the newly added behavior data and the occurrence time of the historical behavior data to which the second object belongs.

In step 204, the electronic device trains the recommendation model based on the propagation loss negatively correlated with a first similarity determined based on a similarity between the influencing feature of each second object and the first feature of each second object.

The trained recommendation model comprises second characteristics of the plurality of objects, and is used for recommending based on the second characteristics of the plurality of objects.

In the embodiment of the disclosure, after two first objects perform an interactive action to obtain new action data, the interactive action can affect a second object related to the first object, so that an influence characteristic representing an influence of the interactive action corresponding to the new action data on the second object can be determined by using a first characteristic of the first object and an interaction time difference of the second object related to the first object, further, since the propagation loss is determined based on a first similarity between the influence characteristic of the second object and the first characteristic of the second object, the recommendation model can learn the influence of the interactive action on the second object in a training process based on the propagation loss, thereby not only training the recommendation model when the interactive action occurs newly, but also combining the influence of the new action data on the second object based on the propagation loss, and further greatly improves the training accuracy of the recommendation model.

In the embodiment of the disclosure, the interaction behavior between the account and the media content is continuously generated, and the new behavior data is continuously generated, so that the electronic device can update and train the recommendation model based on the new behavior data and the historical behavior data influenced by the new behavior data.

Fig. 3 is a flowchart illustrating a recommendation model processing method according to an exemplary embodiment, where an execution subject of the method is an electronic device, and the embodiment of the present disclosure takes training of a recommendation model based on first behavior information as an example, and the method includes the following steps:

in step 301, the electronic device obtains first behavior information, where the first behavior information includes new behavior data and historical behavior data, the new behavior data is used to represent an interaction behavior between two first objects that belong to different types, the types include an account and media content, and the historical behavior data is historical behavior data corresponding to a second object related to any one of the first objects.

For example, one piece of newly added behavior data is (u, v, r, t), where u is an account, v is media content, r is an interaction type corresponding to an interaction behavior, and t is occurrence time of the newly added behavior data. The account may be a user type account or an author type account.

In some embodiments, the first behavior information includes at least two object nodes belonging to different node types and edges connecting between any two object nodes, and the node types include an account type and a media content type; wherein the at least two object nodes include two first object nodes belonging to different types and a second object node directly or indirectly connected to any one of the first object nodes.

The two first object nodes and the first edge connected between the two first object nodes form new behavior data; the first object node and the second object node which belong to different types and the second edge connected between the first object node and the second object node form historical behavior data, and/or any two second object nodes which belong to different types and the third edge connected between any two second object nodes form historical behavior data.

The first behavior information is represented in a graph form, and one piece of behavior data is composed of one edge and two object nodes connected by the edge. The edge type of the edge represents an interaction type of the interaction behavior, such as a type of like or forwarding, and the like. The account types in the node types further include a user type and an author type, and the object nodes include an author account node, a user account node and a media content node. Since the object node has multiple node types and the edge also has multiple edge types, the graph corresponding to the first behavior information is a heterogeneous graph. Further, the user account may perform multiple interactive behaviors on the same media content, such as a praise behavior and a collection behavior, and thus, multiple edges of different edge types may exist between two connected object nodes in the graph, and the graph is a multiple heterogeneous graph. Furthermore, as the new added behavior data is continuously generated, the graph is a dynamic multiple abnormal graph.

In the graph, one account corresponds to at least one of a user node and an author node, and each object node in the graph has a corresponding characteristic, so that one account may have one or two characteristics, that is, at least one of the characteristics of the user node and the characteristics of the author node.

In some embodiments, when the electronic device is a terminal, the terminal obtains new behavior data generated by the terminal, and trains the recommendation model based on the new behavior data and historical behavior data corresponding to the second object. Or, in the case that the electronic device is a server, the server acquires the added behavior data by means of the terminal, and trains the recommendation model based on the added behavior data and the historical behavior data corresponding to the second object. The terminal reports the interactive behavior of the account and the media content of the terminal to the server, and the server generates new behavior data.

In some embodiments, implementations of the electronic device obtaining the first behavior information include: and the electronic equipment samples second behavior information according to the two first objects in the newly added behavior data to obtain the first behavior information, wherein the second behavior information comprises the newly added behavior data and a plurality of pieces of historical behavior data.

Wherein the second behavior information comprises a large amount of behavior data, and the first behavior information is a small amount of behavior data sampled based on the two first objects. And if the second behavior information is represented by the dynamic multiple heterogeneous graph, the first behavior information is an activated subgraph obtained by sampling the dynamic multiple heterogeneous graph. By "activated" is understood that the second object is affected by the newly added behavioural data.

In the embodiment of the disclosure, the first behavior information is obtained by sampling from the second behavior information containing more historical behavior data, so that the recommendation model does not need to be trained based on the second behavior information with larger data volume each time new behavior data is obtained, but only needs to be trained based on the first behavior information with smaller data volume, thereby greatly reducing the data volume of model training, and the historical behavior data in the first behavior information is obtained by sampling according to two first objects in the new behavior data and is historical behavior data related to the new behavior data, and the recommendation model is trained based on the new behavior data and the historical behavior data subsequently, so that the influence of the new behavior data on the historical behavior data can be fully considered, and the accuracy of the recommendation model is improved.

In some embodiments, the implementation manner of the electronic device sampling the second behavior information according to the two first objects in the newly added behavior data to obtain the first behavior information includes: the method comprises the steps that electronic equipment obtains a sampling mode set, wherein the sampling mode set comprises a plurality of sampling modes; determining a sampling mode of each first object from the sampling mode set; for each first object, sampling is carried out in the second behavior information according to the determined sampling mode from the first object, and historical behavior data and newly-added behavior data obtained by sampling form the first behavior information.

In the embodiment of the disclosure, the first behavior information is obtained by sampling from the second behavior information according to the sampling mode corresponding to each first object, so that the first behavior information obtained by sampling is relatively accurate.

Optionally, the second behavior information is represented by a dynamic multi-heterogeneous graph, the sampling modes are collected into a meta-path mode set, the meta-path mode set includes a plurality of meta-path modes, and each meta-path mode indicates one meta-path. Each meta path schema in the meta path schema set may be set as needed, which is not limited by this disclosure. For each first object node connected by the first edge, the electronic device determines at least one meta-path mode of the first object node from the meta-path mode set, and sets the number of paths sampled from the first object node and the path length of each path. The path length corresponds to the number of object nodes included in the path, for example, if the path length is 5, the number of object nodes included in the path is 5.

Correspondingly, for each first object node, the electronic device samples from the dynamic multiple heterogeneous composition corresponding to the second behavior information according to the meta-path mode to obtain a path set corresponding to the first object node, wherein each path comprises at least one edge. Thus, the multiple paths sampled by the two first object nodes constitute the active subgraph.

For example, the meta-path schema set is

The meta path mode is

The path set of the account number node u in the new edge is

The account node u has k paths, each path satisfies a certain mode in the meta-path mode set, and the conditions are as follows:

wherein p is_iThe ith object node, phi (p), representing path p_i) Representing an object node p_iThe type of the node of (a) is,

presentation mode

First, the

Object nodeF (i, L) — (i-1) modL) +1, mod is the remainder operation,

presentation mode

Has a path length of l, ψ (p)_j,p_j+1) Represents an edge (p)_j,p_j+1) Edge type of p_jAnd p_j+1Are respectively an edge (p)_j,p_j+1) Two object nodes that are connected to each other,

presentation mode

First, the

Edge type of individual edge.

It should be noted that, since the path length of each path corresponding to the first object node set in advance may not be equal to the path length of the meta path indicated by the meta path mode, the electronic device may repeat the path in the meta path mode in the sampling process, so that the path length of the meta path mode is long enough.

For example, the electronic device may set the meta-path mode to a symmetric form. Meta-path schema for asymmetric forms in meta-path schema set

n is the number of nodes included in the path, which the electronic device can convert into a meta-path model in a symmetrical form

It should be noted that after the newly added behavior data is obtained and before the recommendation model is trained, the electronic device obtains the first behavior information, so that in the training process of the recommendation model, the first behavior information obtained by sampling in advance can be used in each iteration process of performing the iterative training of the recommendation model based on the first behavior information, and the sampling is not required to be performed again in each iteration training.

In step 302, the electronic device obtains a recommendation model, the recommendation model including first features of a plurality of objects, the plurality of objects including two first objects and at least one second object in historical behavior data.

Alternatively, the first feature can be represented in the form of an embedded vector (embedding).

In some embodiments, the recommended model is an initial model, i.e., an untrained model. Optionally, the first features of the multiple objects included in the recommendation model acquired by the electronic device are features obtained by randomly initializing the recommendation model. The electronic equipment acquires a plurality of pieces of historical behavior data, extracts a plurality of objects from the plurality of pieces of historical behavior data, and randomly initializes the characteristics of the plurality of objects through a recommendation model to obtain first characteristics of the plurality of objects.

In other embodiments, the recommended model is a trained model. Optionally, the electronic device trains the initial model based on a plurality of sample behavior data to obtain the recommended model. The implementation process of training to obtain the recommendation model based on the multiple sample behavior data is shown in fig. 5, and is not described herein again. And if the recommendation model is a trained model, the first features of the plurality of objects included in the recommendation model acquired by the electronic equipment are the first features of the recommendation model which are trained.

In step 303, the electronic device attenuates the first feature of the first object based on the interaction time difference of the first object for each first object to obtain the interaction feature of the first object, where the interaction time difference of the first object is a time difference between the occurrence time of the new added behavior data and the occurrence time of the historical behavior data to which the first object belongs.

The historical behavior data to which the first object belongs comprises at least one piece of historical behavior data. In some embodiments, the historical behavior data to which the first object belongs is historical behavior data corresponding to a latest interaction behavior of the first object. Wherein, the latest means before the interactive behavior corresponding to the newly added behavior data occurs and is closest to the occurrence time of the newly added behavior data. In the embodiment of the disclosure, the interaction time difference of the first object is determined by combining the occurrence time of the interaction behavior performed by the first object at the latest on the basis of the occurrence time of the newly added behavior data, the calculation method is simple, and the efficiency of obtaining the interaction time difference is high.

In other embodiments, the historical behavior data to which the first object belongs includes at least two pieces of historical behavior data, where the at least two pieces of historical behavior data include historical behavior data corresponding to at least two latest interactions performed by the first object. The implementation manner of the electronic device acquiring the interaction time difference of the first object includes: the electronic equipment determines the time difference between the occurrence time of the newly added behavior data and the occurrence time of the historical behavior data corresponding to each interactive behavior in the at least two interactive behaviors, and determines the interactive time difference of the first object based on the at least two time differences.

Wherein at least two interactive activities of the first object are performed sequentially. The at least two interactive behaviors comprise interactive behaviors over a period of time or comprise a certain number of interactive behaviors. The duration of the time period may be set as needed, which is not limited in the embodiment of the present disclosure. Such as a duration of 30 minutes, 1 hour, or 3 hours, etc. The number may be set as desired, and the disclosure is not limited thereto. Such as 5, 10, or 20, etc.

In a possible implementation manner of this embodiment, the determining, by the electronic device, the interaction time difference of the first object based on the at least two time differences includes: the electronic device takes the average of the at least two time differences as the interaction time difference of the first object. In the embodiment of the disclosure, when the interaction time difference of the object is determined, the interaction time corresponding to the multiple interaction behaviors performed by the object at the latest is referred to, so that the accuracy of the determined interaction time difference is higher.

In some embodiments, the electronic device stores each piece of historical behavior data, so that the electronic device can obtain, from the stored historical behavior data, historical behavior data to which the first object belongs. It should be noted that the interaction time difference of the first object does not change with the iterative training of the recommendation model, and then the electronic device may calculate and store the interaction time difference of the first object after acquiring the first behavior information and before training the recommendation model, so that the recommendation model may acquire the stored interaction time difference of the first object in the iterative training process, and thus, repeated calculation is not required, and the training efficiency of the model may be further improved.

In some embodiments, the first feature of the first object includes a first memory feature and a second memory feature. The first memory characteristic is characteristic of the first object in the long term and the second memory characteristic is characteristic of the first object in the short term.

Wherein the first memory characteristics refer to behavior data of the first object generated in a longer period of time, and can characterize the first object from a long time. The second memory characteristic characterizes a recent characteristic of the first object, i.e. the recently generated behavior data has a greater influence on the second memory characteristic. In some embodiments, the first memory feature and the second memory feature in the initial recommendation model are obtained by randomly initializing the recommendation model and cannot accurately represent the feature of the first object, but as the training times of the recommendation model increase, the recommendation model attenuates the second memory feature in combination with newly-added behavior data, so that the second memory feature can be more fit with the recent feature of the first object, and the trained recommendation model includes the more accurate first memory feature and the more accurate second memory feature.

The implementation manner of the electronic device attenuating the first feature of the first object based on the interaction time difference of the first object to obtain the interaction feature of the first object includes the following steps (1) - (2):

(1) the electronic device attenuates the second memory characteristic of the first object based on the interaction time difference of the first object.

In some embodiments, the electronic device, based on the interaction time difference of the first object, implementing the attenuation of the second memory characteristic of the first object comprises: the electronic equipment determines a second attenuation parameter based on the interaction time difference and a learning parameter corresponding to the type of the first object; the second memory characteristic is attenuated based on a second attenuation parameter.

The larger the interaction time difference of the first object is, the longer the time interval between the interaction behavior corresponding to the newly added behavior data and the interaction behavior performed by the first object for the previous time or the previous times is, that is, the frequency of the first object performing the interaction behavior recently is lower, and the second memory feature is difficult to accurately represent the semantic information of the media content recently matched with the first object. That is, for the first object, the second memory characteristic is not greatly different from the first memory characteristic, and therefore, the second attenuation parameter may be set to be smaller so that the second memory characteristic is not greatly different before and after the attenuation. Correspondingly, the smaller the interaction time difference of the first object is, the shorter the time interval between the interaction behavior corresponding to the newly added behavior data and the interaction behavior performed by the first object for the previous time or the previous times is, that is, the higher the frequency of the first object performing the interaction behavior recently is, the second memory feature can accurately represent the semantic information of the media content recently matched with the first object. The second attenuation parameter may be set larger so that the second memory characteristic is distinguishable from the first memory characteristic.

Moreover, the type of the first object comprises a plurality of types of user accounts, author accounts or media contents, and the electronic equipment can determine a proper learning parameter based on the type of the first object so as to obtain a proper second attenuation parameter, so that the attenuation degrees of the first characteristics of different types of first objects are different.

In the embodiment of the disclosure, the interaction time difference of the first object can reflect the time difference between the newly added interaction behavior and the historical interaction behavior of the first object, the time information of the interaction behavior is included, the learning parameter corresponds to the type of the first object, the information of the type of the first object is included, and the second memory characteristic is attenuated by combining the interaction time difference and the learning parameter, so that the second memory characteristic can be attenuated by referring to the time information and the object type information of the first object, and thus the attenuated second memory characteristic better conforms to the recent characteristic of the first object, and the accuracy of the second memory characteristic is improved.

Optionally, the electronic device determines the second attenuation parameter using the following formula:

wherein m is a second attenuation parameter,

u is the first object, phi (u) is the type of the first object, alpha_φ(u)In order to learn the parameters corresponding to the type,

x and y are arbitrary variables, and Δ (u) is the interaction time difference.

After obtaining the second attenuation parameter, the implementation manner of the electronic device attenuating the second memory characteristic based on the second attenuation parameter includes: the electronic device takes the product of the second attenuation parameter and the second memory characteristic as the attenuated second memory characteristic.

(2) And the electronic equipment fuses the first memory characteristics of the first object and the attenuated second memory characteristics to obtain the interaction characteristics of the first object.

In some embodiments, the implementation manner of the fusion is not limited by the embodiments of the present disclosure, and taking summation as an example, the electronic device uses the sum of the first memory feature and the attenuated second memory feature as the interaction feature of the first object.

In the embodiment of the present disclosure, the interaction feature fuses the first memory feature and the second memory feature generated by the behavior data of the first object during the dynamic interaction process, that is, not only the long-term feature of the first object but also the short-term feature attenuated according to the recent occurrence of the interaction behavior are fused, so that the interaction feature can accurately represent the dynamic feature of the first object, and the accuracy is higher.

In the embodiment of the present disclosure, after the interactive feature of the first object is determined, the recommendation model can determine, in combination with the interactive feature of the first object, a model loss of the recommendation model in the iterative training process, so as to perform model training in combination with the model loss. Wherein the model penalty comprises a plurality of types of penalty. Optionally, the electronic device performs the operations of steps 303-305 to determine the interaction loss of the recommendation model, and the electronic device performs the operations of steps 306-308 to determine the propagation loss of the recommendation model. The disclosed embodiment does not impose any limitation on the order of determination of the two model losses.

In step 304, the electronic device determines an attenuation feature of the first object based on the interaction feature of the first object, where the attenuation feature characterizes a feature of the first object after the first feature has been attenuated under the influence of the interaction behavior corresponding to the newly added behavior data.

The attenuation characteristic of the first object is the characteristic of the first object predicted by the recommendation model based on the interaction time difference of the first object in the one-time iterative training process. In some embodiments, the electronic device directly takes the interaction characteristic of the first object as the attenuation characteristic of the first object.

In further embodiments, the first feature of the first object further comprises a context feature, the context feature representing a feature of the first object under the influence of other objects; then the implementation of step 304 includes: the electronic equipment performs weighted fusion on the interaction feature and the context feature of the first object to obtain the attenuation feature of the first object.

The implementation manner of the weighted fusion is not limited in the embodiment of the present disclosure, and taking weighted summation as an example, the electronic device performs weighted summation on the interactive feature and the context feature to obtain the attenuation feature. The weights may be set as desired, for example, the electronic device may directly take the sum of the interaction feature and the context feature as the decay feature.

In the embodiment of the disclosure, the attenuation characteristics of the first object are obtained through weighted fusion, so that the determined attenuation characteristics refer to not only the interactive characteristics of the first object capable of representing dynamic information, but also the context characteristics of the first object under the influence of the interactive behaviors of other objects, and thus, the accuracy of the attenuation characteristics is improved due to the reference of the characteristics of multiple dimensions.

In some embodiments, the context features of the first object include context features of the first object for a plurality of interaction types, and the new behavior data includes a target interaction type corresponding to the interaction behavior. Wherein each interaction type corresponds to a context feature, the electronic device can determine the attenuation feature of the first object based on the context feature of the corresponding type, so as to improve the accuracy of the attenuation feature. In a possible implementation manner of this embodiment, the performing, by the electronic device, weighted fusion on the interaction feature and the context feature of the first object to obtain the attenuation feature of the first object includes: the electronic equipment determines the context characteristics corresponding to the target interaction type from the context characteristics of the first object aiming at the multiple interaction types; and carrying out weighted fusion on the interaction characteristics of the first object and the context characteristics corresponding to the target interaction type to obtain the attenuation characteristics of the first object.

The implementation manner of the weighted fusion is not limited in the embodiment of the present disclosure, and taking weighted summation as an example, the electronic device performs weighted summation on the interaction feature of the first object and the context feature corresponding to the target interaction type to obtain the attenuation feature of the first object. The weight can be set according to the requirement, for example, the electronic device directly takes the sum of the interaction feature and the context feature corresponding to the target interaction type as the attenuation feature.

In the embodiment of the disclosure, the interaction type can represent different interaction behaviors of the object, and the different interaction behaviors can represent the contact between the object and different media contents, taking the object as an account as an example, for example, a click behavior of the account on a short video a and a like behavior of the account on a short video B can represent different preference degrees of the account on the two short videos, and the preference degree of the account on the short video B may be larger. The type of interaction can also have an effect on the characteristics of the object.

In the embodiment of the disclosure, since different interaction types correspond to different context features, the attenuation feature of the first object is determined by the context feature corresponding to the target interaction type based on the newly added behavior data, so that the attenuation feature is more matched with the target interaction type corresponding to the interaction behavior occurring this time, and thus, the accuracy of the attenuation feature is greatly improved.

Optionally, the electronic device determines the attenuation characteristic of the first object using the following formula:

wherein, a is a first object,

in order to be of a attenuating character,

in order to have the first memory feature,

is a second memory characteristic, r is a target interaction type,

the context feature corresponding to the target interaction type, g (x) is a second attenuation parameter,

x and y are arbitrary variables, α_φ(a)To learn the parameters, φ (a) is the type of the first object, Δ_aIs the interaction time difference.

For example, taking the interaction behavior corresponding to the newly added behavior data as an edge (u, v, r, t) in the dynamic multiple anomaly graph, two objects are respectively a connected account node u and a media content node v as an example, and r is an edge classType, t is the interaction time, the attenuation characteristic of the account node u is

The decay characteristic of the media content node v is

In the embodiment of the present disclosure, since the attenuation feature of the first object is obtained by fusing features of multiple dimensions, the accuracy of the attenuation feature is higher.

In step 305, the electronic device determines an interaction loss negatively correlated with a second similarity based on the attenuation characteristics of the two first objects, the second similarity being a similarity between the attenuation characteristics of the two first objects.

Because the features of the account refer to semantic information of the media content matched with the account, and the features of the media content refer to semantic information of the media content, if two first objects perform an interactive behavior, the features of the two first objects should be closer to each other, thereby indicating that the account prefers the media content. In the training process of the recommended model, the training goal of the model is to maximize the similarity between the attenuation features of the two first objects, i.e., the second similarity. Further, the interaction loss is inversely related to the second similarity, and the smaller the interaction loss, the more similar the attenuation characteristics representing the two first objects.

In the embodiment of the disclosure, the interaction loss is negatively correlated with the similarity before the attenuation features of the two first objects, so that the prediction deviation of the recommended model in the training process can be accurately represented by the interaction loss, and the calculation process of the interaction loss is simpler, so that the calculation difficulty of the model loss is reduced.

In some embodiments, the attenuation features are represented in the form of an embedded vector, and the similarity between the attenuation features of the two first objects can be represented by the inner product of the two attenuation features. Optionally, the electronic device determines the interaction loss using the following formula:

wherein the content of the first and second substances,

in order to be a loss of interaction,

z is an arbitrary variable and is a linear variable,

and

respectively the attenuation characteristics of the two first objects.

In the embodiment of the disclosure, the interaction loss of the recommended model is calculated by using the inner product of the two attenuation characteristics, so that the calculation difficulty is reduced, and the similarity between the two attenuation characteristics can be accurately represented, so that the interaction loss can be calculated quickly and accurately.

In step 306, the electronic device determines, for each first object, an influence feature of a second object related to the first object based on an interaction feature of the first object and an interaction time difference of the second object belonging to the same historical behavior data as the first object.

Since any first object may have performed other interactive behaviors before the electronic device acquires the new behavior data, that is, historical behavior data exists, the occurrence of the new behavior data may affect the historical behavior data, and the features of the second object included in the historical behavior data should be updated.

Optionally, the electronic device, based on the interaction characteristic of the first object and the interaction time difference of the second object belonging to the same historical behavior data as the first object, determining the influence characteristic of the second object related to the first object in an implementation manner includes: the electronic equipment determines a first attenuation parameter negatively correlated with the interaction time difference under the condition that the interaction time difference is not larger than a time difference threshold value, and attenuates the interaction characteristic of the first object according to the first attenuation parameter to obtain the influence characteristic of the second object; or, in the case that the interaction time difference is greater than the time difference threshold, determining the preset influence feature as the influence feature of the second object.

Wherein, the time difference threshold is a set time difference. The interaction time difference of the second object is greater than the time difference threshold, which indicates that the first object performs the interaction behavior corresponding to the new behavior data only within a long period of time after performing the interaction behavior corresponding to the historical behavior data, and then the degree of influence of the interaction behavior corresponding to the new behavior data on the second object is small, the electronic device may directly determine the preset influence characteristic as the influence characteristic of the second object, and accordingly, the preset influence characteristic is set to be a small numerical value.

The interaction time difference of the second object is not greater than the time difference threshold, which indicates that the first object performs the interaction behavior corresponding to the new behavior data within a short period of time after performing the interaction behavior corresponding to the historical behavior data, and then the degree of influence of the interaction behavior corresponding to the new behavior data on the second object is large, the electronic device may further determine different first attenuation parameters according to the magnitude of the interaction time difference, so that the influence characteristic of the second object can more accurately indicate the degree of influence on the second object. Optionally, the larger the interaction time difference is, the smaller the representation influence degree is, the smaller the first attenuation parameter is; the smaller the interaction time difference, the greater the degree of influence, and the greater the first attenuation parameter.

The time difference threshold and the preset influence characteristic may be set as needed, which is not limited in the embodiments of the present disclosure, for example, the preset influence characteristic is set to 0, and the time difference threshold is set to 1 hour, 2 hours, or 5 hours.

In the embodiment of the disclosure, because the magnitude of the interaction time difference of the second object can affect the degree of influence of the interaction behavior corresponding to the newly added behavior data on the second object, when the interaction time difference is smaller, it indicates that the occurrence time of the historical behavior data is closer to the occurrence time of the newly added behavior data, and the influence of the newly added behavior data on the historical behavior data is larger, the interaction characteristic of the first object is attenuated according to the attenuation parameter negatively related to the interaction time difference, the influence of the newly added behavior data can be better reflected, and thus the accuracy of the influence characteristic is improved. When the interaction time difference is large, the occurrence time of the historical behavior data and the newly added behavior data is far away, and the influence of the newly added behavior data on the historical behavior data is small, so that the preset influence characteristic is determined to be the influence characteristic of the second object, the interaction characteristic of the first object does not need to be attenuated according to the attenuation parameter which is negatively related to the interaction time difference, and the calculated amount is saved.

Optionally, the electronic device determines the influence characteristic of the second object using the following formula:

wherein p is historical behavior data of the same genus of the first object and the second object, and u is_pIs a first object, b_pIs the second object and is the first object,

is an interactive feature of the first object,

is an influence characteristic of the second object, Δ (t)_p)＝t-t_pT is the occurrence time of the newly added behavior data, t_pIs the time of occurrence of the historical behavioral data,

w is an arbitrary variable, τ is a threshold,

is a first attenuation parameter.

In the embodiment of the disclosure, the interaction feature of the attenuated first object is used as the influence feature of the second object, so that the feature of the second object under the influence of the interaction behavior corresponding to the newly-added behavior data is accurately represented, and the accuracy of the influence feature is improved.

In step 307, the electronic device continues to determine the influence characteristic of another second object based on the influence characteristic of the second object and the interaction time difference of another second object belonging to the same historical behavior data with the second object, until the influence characteristic of each second object related to the first object in the first behavior information is determined.

For a second object related to the first object, the degree of influence of the first object on the second object decreases with the distance between the second object and the first object, that is, the greater the distance between the second object and the first object, the smaller the degree of influence of the second object on the first object, and the less information the first object transmits to the second object. Taking an object as an object node in the dynamic multiple heterogeneous graph as an example, the distance between two objects can be represented by the number of edges between two object nodes, for example, if two object nodes are connected by one edge, the two object nodes are closer to each other, and if two object nodes are connected by other object nodes and multiple edges, the two object nodes are farther from each other.

In some embodiments, the electronic device determines, based on the influence characteristic of the second object and an interaction time difference between the second object and another second object belonging to the same historical behavior data as the second object, that an implementation manner of the influence characteristic of the another second object is the same as that of step 306, and details are not repeated here. Step 303 and steps 306-307 are one implementation of the electronic device determining, for each first object, an influence feature of a second object based on a difference in interaction time between a first feature of the first object and the second object associated with the first object.

In the embodiment of the present disclosure, the interaction behavior performed by two first objects affects the second object, and the farther the time difference between the interaction behavior and the interaction behavior performed by the second object is, the smaller the influence is, and the influence characteristic of the second object is determined based on the interaction characteristic of the first object and the interaction time difference of the second object, so that the influence characteristic is determined based on the interaction characteristic with reference to the interaction time difference, and thus the accuracy of the influence characteristic is determined to be higher. And the first behavior information contains a plurality of objects, and the first object can be directly or indirectly related to a plurality of second objects, so that according to the incidence relation between the first object and the plurality of second objects, the influence characteristic of the second object directly related to the first object is determined firstly, and then the influence characteristic of another second object directly related to the second object is determined.

In some embodiments, the electronic device may be configured to determine, for each first object, an influence feature of a second object based on a difference in interaction time between a first feature of the first object and the second object associated with the first object, in a graph structure, and accordingly, the process may include the steps of:

the electronic equipment attenuates the first characteristics of the first object nodes based on the interaction time difference of the first object nodes to obtain the interaction characteristics of the first object nodes, wherein the interaction time difference of the first object nodes is the time difference between the occurrence time of a first edge and the occurrence time of a second edge connected with the first object nodes; determining an influence characteristic of a second object node based on an interaction characteristic of the first object node and an interaction time difference of the second object node directly connected with the first object node; and continuing to determine the influence characteristics of the other second object node based on the interaction time difference between the influence characteristics of the second object node and the other second object node directly connected with the second object node until the influence characteristics of each second object node directly or indirectly connected with any first object node in the first behavior information are determined.

And for each first object node of the first edge, the interactive characteristics of the first object node are propagated from the first object node along the path in the active subgraph in turn, so that the interactive characteristics of the first object node can be propagated to each second object node on the path, and the interactive characteristics are gradually attenuated in the propagation process.

In the embodiment of the disclosure, the association relationship between the plurality of objects is clearer by representing the behavior data in the form of a graph structure, so according to the association relationship between the first object node and the plurality of second object nodes, the influence characteristic of the second object directly connected with the first object node is determined first, then the influence characteristic of another second object node directly connected with the second object node is determined, and the manner of sequentially determining the influence characteristic of each second object node through the one-by-one processing can more closely reflect the strength of the association between the first object and each second object, thereby improving the accuracy of the influence characteristic of each second object.

In step 308, the electronic device determines a propagation loss that is negatively correlated with a first similarity determined based on a similarity between the influencing feature of each second object and the first feature of each second object.

Wherein the first feature of the second object comprises a context feature, and the context feature characterizes a feature of the second object under the influence of other objects; implementations of the electronic device determining the first similarity include: the electronic equipment determines the similarity between the influence characteristic and the context characteristic of each second object; and fusing the similarity of at least one second object to obtain a first similarity.

The embodiment of the present disclosure does not limit the implementation manner of the fusion, and takes summation as an example, and the electronic device takes the sum of the similarities of at least one second object as the first similarity.

In the embodiment of the disclosure, since the context feature of the second object can represent the feature of the second object under the influence of other objects, and the influence feature of the second object is the feature of the second object under the influence of the interaction behavior corresponding to the newly-added behavior data, the similarity between the influence feature of the second object and the context feature can embody the prediction accuracy of the recommendation model on the influence feature of the second object, so that the determined first similarity is more accurate, and the accuracy of the propagation loss is further improved.

The training goal of the recommendation model training should be to maximize the similarity between the impact feature and the context feature of the second object, that is, the greater the similarity, the smaller the propagation loss of the recommendation model, and the negative correlation between the first similarity and the propagation loss of the recommendation model. In some embodiments, the features are represented in the form of an embedded vector, and the similarity between two features can be represented by the inner product of the two features, the electronic device can determine the propagation loss using the following formula:

wherein the content of the first and second substances,

for propagation loss, p_u∪p_vIs the first behavior information, p_uHistorical behavior data, p, corresponding to a second object associated with a first object u_vHistorical behavior data corresponding to a second object related to the first object v, p is any piece of historical behavior data in the first behavior information, r is an interaction type to which an interaction behavior corresponding to the historical behavior data p belongs,

is the context feature corresponding to the interaction type r of the second object,

is an influencing feature of the second object.

In the embodiment of the disclosure, the propagation loss of the model is calculated by using the inner product of the influence feature and the context feature, so that the calculation difficulty is reduced, the similarity of the two features can be accurately represented, and the propagation loss can be calculated quickly and accurately.

For example, taking the first behavior information as the activation subgraph as an example, the propagation loss of the recommendation model can be determined by using the following formula:

wherein the account number node is u, the media content node is v,

in order to activate the sub-graph,

is a path set corresponding to the account node u,

a set of paths corresponding to media content node v, p is any path in the active subgraph, i is the sequence number of the edge on path p,<v_i,r_i>indicates the type of the edge is r_iIs directed to node v_iThe propagation of the beam is carried out,

is a node v_iThe characteristics of the influence along the path p,

is a node v_iEdge-to-edge type r_iThe context vector of (1), χ (·) is an indicative function. The indicative function being represented at the event

In the case of a true event,

get 1 at event

In the case of a false event,

take 0.

In step 309, the electronic device fuses the propagation loss and the interaction loss to obtain a model loss of the recommended model.

The embodiment of the present disclosure does not limit the implementation manner of the fusion, and takes summation as an example, and the electronic device uses the sum of the propagation loss and the interaction loss as the model loss.

In some embodiments, the model penalty further comprises a negative sampling penalty. The negative sampling loss represents the loss obtained by training the recommendation model according to the negative sample, and the interaction loss and the propagation loss in the above case are the losses obtained by training the recommendation model according to the positive sample. The positive example is also real behavior data, and the negative example is false behavior data, that is, at least two objects in each piece of behavior data in the negative example have no interactive behavior.

Optionally, the implementation manner of step 309 includes: and the electronic equipment fuses the interaction loss, the propagation loss and the negative sampling loss to obtain the model loss of the recommended model. The electronic device determines the model loss using the following equation:

wherein the content of the first and second substances,

in order to be a loss of interaction,

in order to achieve a propagation loss,

is a negative sampling loss.

In some embodiments, implementations of the electronic device determining negative sampling loss include: the electronic equipment determines a third object corresponding to each first object from second behavior information, wherein the second behavior information comprises newly added behavior data and a plurality of pieces of historical behavior data, and the third object is other objects except the first object in the second behavior information; a negative sampling loss negatively correlated to a third similarity is determined, the third similarity being determined based on a similarity between the interaction feature of the first object and a context feature of a third object, the context feature characterizing features of the third object under influence of other objects.

In some embodiments, the third object is a non-interactive object of the first object, that is, the third object and the first object do not belong to any piece of historical behavior data at the same time. Taking the first object as an account as an example, the non-interactive object of the account may be an account or media content, and since the account is not interacted with the non-interactive object, it indicates that the user to which the account belongs has little interest in the media resource corresponding to the non-interactive object, and the similarity between the interactive feature of the account and the context feature of the non-interactive object is small. In the training process of the recommendation model, the training goal of the recommendation model should be to minimize the similarity between the interaction features of the first object and the context features of the third object.

In the embodiment of the disclosure, the second behavior information is negatively sampled to provide a negative sample for the recommendation model, and the negative sampling loss of the recommendation model is determined, so that the negative sampling loss is additionally added on the basis of propagation loss and interaction loss, the recommendation model can be trained by combining the positive sample and the negative sample, and the training accuracy of the recommendation model is further improved.

In some embodiments, the electronic device determines, from the second behavior information, an implementation manner of the third object of each of the first objects, including: and the electronic equipment determines a third object of each first object from the second behavior information based on the mode of the target probability distribution. The target probability distribution can be set according to the requirement, such as random distribution, uniform distribution or gaussian distribution. For each first object, the electronic device determines the probability of each object except for the two first objects in the second behavior information based on the target probability distribution, determines a random number, the value range of the random number is (0, 1), and takes the object to which the probability corresponding to the range of the random number belongs as the third object. The electronic device may set the number of the third objects, so as to select the number of the third objects from the second behavior information in sequence according to the implementation manner.

In some embodiments, the features are represented in the form of an embedded vector, and the similarity between two features can be represented by the inner product of the two features, the electronics can determine a negative sampling loss using the following equation:

wherein n is_sIs the number of third objects, s represents a negative sample, j is 1, 2, 3, … …, n_s，P_NegFor a target probability distribution that is satisfied by negative sampling, k is the probability that the third object corresponds under the target probability distribution, E denotes expectation, q1 is the third object of the first object u, q2 is the third object of the first object v,

is a contextual feature of the third object q1 for the edge type r,

is an interactive feature of the first object u,

is a contextual feature of the third object q2 for the edge type r,

is an interactive feature of the first object v.

In the embodiment of the disclosure, the negative sampling loss of the recommendation model is calculated by using the inner product of the interactive feature of the first object and the context feature of the third object, so that the calculation difficulty is reduced, the similarity of the two features can be accurately represented, and the negative sampling loss can be calculated quickly and accurately.

In the embodiment of the present disclosure, the recommended model further includes model parameters, and the electronic device performs the operation of step 310 after obtaining the model loss.

In step 310, the electronic device updates the model parameters in the recommendation model and the first features of the plurality of objects based on the model loss to obtain a trained recommendation model, where the trained recommendation model includes the second features of the plurality of objects, and the trained recommendation model is used for recommending based on the second features of the plurality of objects.

In some embodiments, the electronic device inputs new added behavior data into a recommended model determined in an ith-1 iteration process in an ith iteration process of model training to obtain a model loss of the ith iteration process, and updates model parameters determined in the ith-1 iteration process and first characteristics of a plurality of objects based on the model loss, wherein i is a positive integer greater than 1; and (3) performing an (i + 1) th iteration process based on the updated model parameters and the first characteristics of the plurality of objects, and repeating the iteration process of the training until the training meets the target condition.

In some embodiments, the target condition met by the training is that the number of training iterations of the model reaches a target number, which is a preset number of training iterations, such as 1000; alternatively, the training satisfies a target condition that the model loss satisfies a target threshold condition, such as the model loss is less than 0.00001. The embodiments of the present disclosure do not limit the setting of the target conditions.

In the embodiment of the disclosure, the propagation loss and the interaction loss are fused to obtain the model loss, and the model loss can reflect the loss caused by the influence of the newly added behavior data on the second object and the loss caused by the influence of the interaction behavior between the two first objects on the first object, so that the recommendation model is iteratively trained based on the model loss, so that the better model parameters and the second characteristics of the objects can be trained, the recommendation model with better prediction capability is obtained, and the prediction accuracy of the recommendation model is further improved.

In the embodiment of the present disclosure, the electronic device may directly use the interaction loss as a model loss, and train the recommended model based on the interaction loss, that is, after the electronic device performs steps 301 to 305, the electronic device directly performs the operation of step 310. Alternatively, the electronic device may directly use the propagation loss as a model loss, and train the recommended model based on the propagation loss, that is, after the electronic device performs steps 301 to 303, steps 306 to 308 are performed, and then the operation of step 310 is performed. Or, the electronic device may fuse the interaction loss and the propagation loss, and train the recommended model according to the model loss obtained by fusing, that is, after the electronic device completes 301 to 308, the electronic device performs 309 to 310 operations.

In the embodiment of the disclosure, the recommendation model is trained by combining the interaction loss determined based on the similarity between the attenuation features of the two first objects and the propagation loss determined based on the similarity between the influence feature of the second object and the own first feature, so that the recommendation model can refer to the two similarities in the training process, thereby improving the training accuracy.

It should be noted that the electronic device can also directly train the recommendation model based on the propagation loss and the negative sampling loss, and then the implementation manner of the electronic device training the recommendation model based on the propagation loss negatively correlated to the first similarity includes: the electronic device trains the recommendation model based on the propagation loss and the negative sampling loss that are negatively correlated with the first similarity. Optionally, the implementation manner of this process is the same as that of steps 309 to 310, and is not described herein again.

In the embodiment of the disclosure, since the negative sampling loss is determined based on the similarity before the second behavior information is negatively sampled to obtain the first feature of the third object and the interactive feature of the first object, and the propagation loss is determined based on the similarity between the influence feature of the second object and the first feature of the second object, the recommendation model is trained by combining the negative sampling loss and the propagation loss, so that the recommendation model can refer to the two similarities in the training process, thereby improving the training accuracy.

It should be noted that, in the embodiment of the present disclosure, the number of the added behavior data is 1 as an example for explanation. In some embodiments, each time the electronic device acquires one piece of new behavior data, the electronic device stores the new behavior data, and when the number of the stored new behavior data reaches a certain number, the electronic device acquires first behavior information based on the stored pieces of new behavior data, thereby training the recommendation model. In each iteration process of the model, for each newly added behavior data, the electronic device determines the model loss according to the above-mentioned manner in steps 301 to 309, that is, the electronic device sequentially inputs the first behavior information corresponding to each newly added behavior data into the recommended model to obtain the model loss. And the electronic equipment fuses the obtained multiple model losses to obtain the model loss of the iteration process. For example, the electronic device determines the sum of the model losses as the model loss of the current iteration.

The recommendation model processing method provided by the embodiment of the disclosure can effectively model multiple heterogeneous graphs and streaming dynamic heterogeneous graphs, that is, effectively model dynamically added behavior data with various object types. In addition, different from training of some recommendation models in the related art in an inward aggregation manner, the recommendation model provided by the embodiment of the disclosure is trained based on a sampling update propagation architecture, so that noise influence caused by drastic change of neighbor nodes in behavior data is avoided. Further, in the modeling stage, for the newly added behavior data, the recommendation model samples the first behavior information affected by the newly added behavior data according to a specified sampling mode set, so that the interaction characteristics of the two first objects are determined according to the time information and the object types, the interaction characteristics are propagated to the second object in the first behavior information, and the influence characteristics of the second object are determined according to the time information and the interaction types.

In addition, the recommendation model can be trained only based on newly-added behavior data which is dynamically generated, so that the recommendation model can be updated in real time according to the newly-added behavior data in an online environment. It should be noted that the real-time updating includes performing update training on the recommended model each time a new behavior data is generated, and also includes performing update training on the recommended model according to a certain amount of new behavior data when the new behavior data reaches a certain amount.

For example, fig. 4 is a schematic diagram of a training process of a recommendation model according to an exemplary embodiment, where the second behavior information is a dynamic multiple heterogeneous graph, the interaction behavior corresponding to the behavior data is an edge in the dynamic multiple heterogeneous graph, the object is an object node, and the recommendation model is composed of three modules, which are respectively: an Active Graph Sampling module (Active Graph Sampling), a relationship-specific Update module (relationship-specific Update), and a Time-aware Update module (Time-aware prediction). The active graph sampling module samples the dynamic multiple heterogeneous graph to obtain an active sub-graph, and the attenuation characteristics of each first object node in the first edge are determined through the relationship perception updating module, so that the L is determined₁Determining L through determining the influence characteristics of each second object node in the activation subgraph by a time perception updating module so as to determine Interaction Loss (Interaction Loss)₂Propagation Loss (Propagation Loss).

The method comprises the following steps that a is an Author (Author) Node, u is a User (User) Node, v is a media content (Video) Node, t is occurrence time, interaction types comprise a Click type (Click), a Like type (Like), a Forward type (Forward) and a release type (Upload), the interaction type of a first edge (A New Click) is a Click type, two object nodes of the first edge are interaction nodes, other object nodes in an activation subgraph are affected second object nodes (infected nodes), the characteristics of the object nodes comprise a first Memory characteristic, namely a Long-term Memory characteristic (Long-term Memory), a second Memory characteristic, namely a Short-term Memory characteristic (Short-term Memory), and a Context characteristic (relationship-specific Context), which is limited by the edge type, and g represents a second attenuation parameter.

Compared with other STOA (State of the art currently most advanced) recommendation models, the SUPA (Sampling update propagation architecture) recommendation model trained by the recommendation model processing method provided by the embodiment of the disclosure has a better recommendation effect on a plurality of data sets, and compared with a suboptimal model, when recommending media contents to an account, under the condition of recalling 50 media contents, an MRR (Mean Recommal Rank, ranking of correct recall value in recall record) index has a relative promotion of 23.62% on average. By the SUPA recommendation model provided in the embodiment of the present disclosure, different influences on object features caused by different object types and different interaction types changing over time are considered, semantic information and time information in behavior data are better utilized, rich semantic information in dynamic streaming data can be better captured, and differential expressions of the features of the object under different interaction behaviors at different times are learned.

While the embodiment shown in fig. 3 described above describes the training process of the recommendation model, in another embodiment, before training the recommendation model based on the first behavior information, the electronic device can train the recommendation model based on the sample behavior data to obtain a recommendation model with a certain recommendation accuracy.

Fig. 5 is a flowchart illustrating a recommendation model processing method according to an exemplary embodiment, and referring to fig. 5, the execution subject of the method is an electronic device, including the steps of:

in step 501, the electronic device obtains a plurality of pieces of sample behavior data, each piece of sample behavior data representing an interaction behavior between two sample objects belonging to different types.

The sample object is an account or media content. In some embodiments, the sample behavior data is behavior data generated by interaction between the account and the media content during the historical interaction process. That is, the sample behavior data may be historical behavior data in the second behavior information, except for the newly added behavior data.

In step 502, the electronic device divides the plurality of pieces of sample behavior data into a plurality of sample sets according to the sequence of the occurrence time from far to near, and the number of the sample behavior data contained in each sample set is the same.

Wherein, a plurality of pieces of sample behavior data can be input into the recommendation model in the form of data tables (edges), and the occurrence time of each piece of sample behavior data can be represented in the form of a time stamp. In some embodiments, the electronic device trains the recommendation model (model) in batches due to the large number of pieces of sample behavior data. And since the plurality of sample behaviors are sorted according to the sequence of the interaction time from far to near, the recommendation model can refer to the influence of the occurrence time of the behavior data on the determination of the object characteristics, and each adjacent target quantity (batch _ size) piece of the behavior data is used as a sample set (batch). The present disclosure does not impose limitations on the setting of the target number. For example, the number of pieces of sample behavior data is 1000, and a sample set may be formed for every 100 pieces of sample behavior data.

In step 503, the electronic device trains recommendation models in turn based on the plurality of sample sets.

In some embodiments, the electronic device iteratively trains the recommendation model within each sample set. Each iteration target turn (valid _ interval), the effect of the recommendation model is verified based on a verification set, which may be a certain number (valid _ size) of sample behavior data later in the interaction time in the current sample set. The training termination condition for each sample set, that is, the target condition that the training satisfies, may be that the effect of the recommended model is continuously increased but reaches the maximum number of training iterations (max _ iter), or may also be that the number of successive descending rounds of the effect of the recommended model exceeds the maximum value (max _ probability).

After the iterative training of each sample set is terminated, selecting a recommended model with the best verification effect to continue training the next sample set until the training of a plurality of sample sets is completed. In some embodiments, for each sample set, the implementation of step 503 is the same as the implementation of steps 301 to 310, and is not described herein again.

It should be noted that the embodiment of the present disclosure takes a dynamic multiple heterogeneous graph as an example for description, and in addition, the electronic device may also train behavior data in the form of the same graph or a static graph through the recommendation model provided by the embodiment of the present disclosure. The isomorphic graph is a graph in which the number of node types and the number of edge types are both 1. The static map is a map in which the occurrence time of each edge is the same. Accordingly, when training the behavior data in the form of the same composition, the electronic device sets both the interaction type and the object type to 1. When the static graph type behavior data is trained, the electronic equipment sets the occurrence time of each piece of behavior data to be the same time.

In the embodiment of the present disclosure, since the interaction behavior between the account and the media content is continuously generated, the amount of behavior data included in the second behavior information is continuously increased. The initial model is trained according to the existing behavior data to obtain a recommendation model with high recommendation accuracy, and the recommendation model is updated and trained according to the newly added behavior data, so that the recommendation model is trained by adopting a single-pass traversal training frame, and the recommendation model does not need to be updated and trained by using all historical behavior data and the newly added behavior data after the newly added behavior data is obtained every time, thereby greatly improving the model training efficiency.

In some embodiments, after the recommendation model is trained, the electronic device can call the recommendation model to make a recommendation. Fig. 6 is a flowchart illustrating a media content recommendation method according to an exemplary embodiment, and referring to fig. 6, the method is executed by an electronic device and includes the following steps:

in step 601, the electronic device obtains a recommendation model, where the recommendation model includes object characteristics of a plurality of objects, and the plurality of objects include a user account and media content.

In step 602, the electronic device determines, based on account characteristics of a first user account and content characteristics of first media content, a degree of similarity between the first user account and the first media content before the first user account is determined, where the first user account is any one of the user accounts, and the first media content is any one of the media content.

In step 603, the electronic device determines to recommend the first media content to the first user account based on the similarity.

The plurality of objects comprise a first object and a second object, the first object is an object in the newly added behavior data, the second object is an object related to the first object, the recommendation model is obtained based on propagation loss training which is negatively related to first similarity, the first similarity is determined based on similarity between influence characteristics of the second object and object characteristics of the second object, and the influence characteristics represent characteristics of the second object under the influence of the interaction behavior corresponding to the newly added behavior data.

In some embodiments, the recommendation model includes at least one of a characteristic corresponding to a user type and a characteristic corresponding to an author type of the account. The recommendation model can recommend media content for an account of a user type, namely a user account, and when the account of the account is of the user type, the electronic equipment acquires characteristics of the account user type; when the account number has two types, namely an author type and a user type, the electronic equipment also acquires the characteristics of the account number user type.

Fig. 7 is a flowchart illustrating a method for recommending media content according to an exemplary embodiment, where an execution subject of the method is an electronic device, and the embodiment of the present disclosure takes an example in which the electronic device recommends media content to an account, and the method includes the following steps:

in step 701, the electronic device obtains a recommendation model, where the recommendation model includes object characteristics of a plurality of objects, and the plurality of objects include a user account and media content.

Wherein, the recommendation model is obtained based on the training of step 301 to step 310, which is not described herein again. In the embodiment of the present disclosure, media content is recommended to an account of a user type, that is, a user account.

In step 702, the electronic device determines a fusion characteristic of the first user account based on the account characteristic of the first user account, and determines a fusion characteristic of the first media content based on the content characteristic of the first media content.

The first user account is any one of the user accounts, and the first media content is any one of the media contents. In some embodiments, the account characteristics include a first memory characteristic and a second memory characteristic of the first user account, the first memory characteristic characterizing long-term characteristics of the first user account, the second memory characteristic characterizing short-term characteristics of the first user account; the content characteristics include a first remembered characteristic and a second remembered characteristic of the first media content, the first remembered characteristic of the first media content characterizing long-term characteristics of the first media content, the second remembered characteristic of the first media content characterizing short-term characteristics of the first media content.

Optionally, the determining, by the electronic device, an implementation manner of the fusion feature of the first user account based on the account feature of the first user account includes: the electronic equipment fuses the first memory characteristics and the second memory characteristics of the first user account to obtain the fusion characteristics of the first user account. The electronic equipment determines the implementation mode of the fusion characteristic of the first media content based on the content characteristic of the first media content, and the implementation mode comprises the following steps: the electronic equipment fuses the first memory characteristics and the second memory characteristics of the first media content to obtain the fusion characteristics of the first media content.

The embodiment of the present disclosure does not limit the implementation manner of the fusion, and takes summation as an example, the electronic device uses a sum of the first memory characteristic and the second memory characteristic of the first user account as the fusion characteristic of the first user account, and uses a sum of the first memory characteristic and the second memory characteristic of the first media content as the fusion characteristic of the first media content.

In other embodiments, the account characteristics include a contextual characteristic of the first user account, the contextual characteristic representing a characteristic of the first user account under the influence of other objects; the implementation manner of the electronic device fusing the first memory feature and the second memory feature of the first user account to obtain the fused feature of the first user account includes: the electronic equipment fuses the first memory feature, the second memory feature and the context feature of the first user account to obtain a fusion feature of the first user account. The process of determining the fusion characteristics of the first media content is the same, and is not described herein again.

It should be noted that the context features of the first user account include context features of the first user account for multiple interaction types, and the electronic device may determine a fusion feature of the first user account for each interaction type. Similarly, the electronic device may also determine a fusion characteristic of the first media content for each interaction type.

In the embodiment of the disclosure, the fusion feature of the first user account is obtained by fusing the features of multiple dimensions of the first user account, so that the fusion feature not only refers to the memory feature corresponding to the interaction behavior performed by the first user account, but also refers to the context feature corresponding to the interaction behavior performed by the first user account under the influence of other objects, thereby improving the accuracy of the fusion feature.

In step 703, the electronic device determines a similarity between the fused feature of the first user account and the fused feature of the first media content.

The similarity may be cosine similarity or other parameters that can measure the similarity between features. In the embodiment of the present disclosure, the description will be given by taking the cosine similarity as an example. In some embodiments, the electronic device determines a similarity of the first user account and the first media content for the same interaction type if the fused feature includes fused features for multiple interaction types. For example, the plurality of interaction types include a like type, a forward type, and a collection type, and for each interaction type, the electronic device determines a similarity.

In some embodiments, steps 702-703 are one implementation of the electronic device determining a similarity between the first user account and the first media content based on account characteristics of the first user account and content characteristics of the first media content.

In the embodiment of the disclosure, the fusion feature of the object fuses the first memory feature and the second memory feature generated by the behavior data of the object through the fusion feature of the object during the dynamic interaction process, that is, not only the long-term feature of the object but also the short-term feature attenuated according to the recent occurrence of the interaction behavior, so that the fusion feature can accurately represent the dynamic feature of the object, so that the change condition of the feature under the influence of the dynamically added behavior data of the first user account and the first media content can be respectively referred to based on the similarity determined by the fusion feature of the first user account and the fusion feature of the first media content, thereby improving the accuracy of the similarity.

In step 704, the electronic device determines to recommend the first media content to the first user account based on the similarity.

In some embodiments, the electronic device recommends the first media content to the first user account with a similarity greater than a similarity threshold. Or the electronic equipment recommends a certain number of first media contents with the top-ranked similarity to the first user account. The number may be set as needed, and the embodiment of the present disclosure does not limit this.

For example, taking the heteromorphic graph as an example, for a user account node in the graph, by calculating the similarity between the user account node and a media content node in the graph, an edge that may be established in the future by the user account node can be determined.

FIG. 8 is a block diagram illustrating a media content recommender, according to an exemplary embodiment. Referring to fig. 8, the apparatus includes:

a first obtaining unit 801 configured to perform obtaining of a recommendation model, where the recommendation model includes object characteristics of a plurality of objects, and the plurality of objects include a user account and media content;

a first determining unit 802, configured to perform determining, based on account characteristics of a first user account and content characteristics of first media content, a similarity between the first user account and the first media content, where the first user account is any one of the user accounts, and the first media content is any one of the media content;

a recommending unit 803 configured to perform recommending the first media content to the first user account based on the similarity determination;

In some embodiments, the account characteristics include a first memory characteristic and a second memory characteristic of the first user account, the first memory characteristic characterizing long-term characteristics of the first user account, the second memory characteristic characterizing short-term characteristics of the first user account; the content characteristics comprise a first memory characteristic and a second memory characteristic of the first media content, the first memory characteristic of the first media content is characterized by the long-term characteristics of the first media content, and the second memory characteristic of the first media content is characterized by the short-term characteristics of the first media content;

a first determining unit 802, configured to perform fusion of a first memory feature and a second memory feature of a first user account to obtain a fusion feature of the first user account, and perform fusion of the first memory feature and the second memory feature of the first media content to obtain a fusion feature of the first media content; a similarity between the fusion characteristics of the first user account and the fusion characteristics of the first media content is determined.

the first determining unit 802 is configured to perform fusion of the first memory feature, the second memory feature, and the context feature of the first user account to obtain a fusion feature of the first user account.

Fig. 9 is a block diagram illustrating a recommendation model processing apparatus according to an example embodiment. Referring to fig. 9, the apparatus includes:

a second obtaining unit 901, configured to perform obtaining of first behavior information, where the first behavior information includes new behavior data and historical behavior data, the new behavior data is used to represent an interaction behavior between two first objects belonging to different types, the types include an account and media content, and the historical behavior data is historical behavior data corresponding to a second object related to any one of the first objects;

a third obtaining unit 902 configured to perform obtaining a recommendation model, the recommendation model including first features of a plurality of objects, the plurality of objects including two first objects and at least one second object in historical behavior data;

a second determining unit 903, configured to perform, for each first object, determining influence features of the second object based on an interaction time difference between a first feature of the first object and a second object related to the first object, where the influence features of any second object characterize a feature of the second object under the influence of an interaction behavior corresponding to the new behavior data, and the interaction time difference of the second object is a time difference between an occurrence time of the new behavior data and an occurrence time of historical behavior data to which the second object belongs;

a training unit 904 configured to perform training of the recommendation model based on propagation loss negatively correlated to a first similarity determined based on a similarity between the impact feature of each second object and the first feature of each second object, the trained recommendation model comprising second features of the plurality of objects, and the trained recommendation model for recommending based on the second features of the plurality of objects.

In some embodiments, the second determining unit 903 is configured to perform, for each first object, attenuating the first feature of the first object based on an interaction time difference of the first object to obtain an interaction feature of the first object, where the interaction time difference of the first object is a time difference between an occurrence time of the new added behavior data and an occurrence time of the historical behavior data to which the first object belongs;

determining an influence characteristic of a second object related to the first object based on an interaction characteristic of the first object and an interaction time difference of the second object belonging to the same historical behavior data with the first object;

and continuing to determine the influence characteristics of the other second object based on the influence characteristics of the second object and the interaction time difference of the other second object belonging to the same historical behavior data with the second object until the influence characteristics of each second object related to the first object in the first behavior information are determined.

In some embodiments, the second determining unit 903 is configured to perform, in a case that the interaction time difference is not greater than the time difference threshold, determining a first attenuation parameter negatively correlated to the interaction time difference, and attenuating the interaction feature of the first object according to the first attenuation parameter to obtain an influence feature of the second object; alternatively, the first and second electrodes may be,

a second determination unit configured to perform determining the preset influence feature as an influence feature of the second object if the interaction time difference is greater than the time difference threshold.

the two first object nodes and the first edge connected between the two first object nodes form new behavior data;

the first object node and the second object node which belong to different types and the second edge connected between the first object node and the second object node form historical behavior data, and/or any two second object nodes which belong to different types and the third edge connected between any two second object nodes form historical behavior data;

a second determining unit 903 configured to perform, for each first object node, attenuating the first feature of the first object node based on an interaction time difference of the first object node, to obtain an interaction feature of the first object node, where the interaction time difference of the first object node is a time difference between an occurrence time of the first edge and an occurrence time of a second edge to which the first object node is connected;

determining an influence characteristic of a second object node based on an interaction characteristic of the first object node and an interaction time difference of the second object node directly connected with the first object node;

and continuing to determine the influence characteristics of the other second object node based on the interaction time difference between the influence characteristics of the second object node and the other second object node directly connected with the second object node until the influence characteristics of each second object node directly or indirectly connected with any first object node in the first behavior information are determined.

In some embodiments, the apparatus further comprises:

the third determining unit is configured to perform attenuation on the first characteristics of the first objects based on the interaction time difference of the first objects for each first object to obtain attenuation characteristics, the attenuation characteristics represent characteristics of the first objects after the first characteristics of the first objects are attenuated under the influence of the interaction behaviors corresponding to the newly added behavior data, and the interaction time difference of the first objects is the time difference between the occurrence time of the newly added behavior data and the occurrence time of the historical behavior data to which the first objects belong;

determining an interaction loss negatively correlated with a second similarity based on the attenuation features of the two first objects, the second similarity being the similarity between the attenuation features of the two first objects;

a training unit configured to perform training of the recommendation model based on the propagation loss and the interaction loss negatively correlated with the first similarity.

In some embodiments, the first feature of the first object comprises a first memory feature characterizing long-term features of the first object and a second memory feature characterizing short-term features of the first object; a third determination unit configured to perform an attenuation of a second memory characteristic of the first object based on the interaction time difference of the first object; and fusing the first memory characteristics of the first object and the attenuated second memory characteristics to obtain the attenuation characteristics of the first object.

In some embodiments, the third determining unit is configured to perform determining the second attenuation parameter based on the interaction time difference of the first object and the learning parameter corresponding to the type of the first object; the second memory characteristic is attenuated based on a second attenuation parameter.

In some embodiments, the first feature of the first object further comprises a contextual feature, the contextual feature characterizing features of the first object under the influence of other objects; and the third determining unit is configured to perform weighted fusion on the first memory characteristics, the attenuated second memory characteristics and the context characteristics of the first object to obtain the attenuation characteristics of the first object.

In some embodiments, the context features of the first object include context features of the first object for a plurality of interaction types, and the newly added behavior data includes a target interaction type corresponding to the interaction behavior;

the third determining unit is configured to determine the context characteristics corresponding to the target interaction type from the context characteristics of the first object aiming at the multiple interaction types; and performing weighted fusion on the first memory characteristics of the first object, the attenuated second memory characteristics and the context characteristics corresponding to the target interaction type to obtain the attenuation characteristics of the first object.

In some embodiments, the recommended model further includes model parameters, and the training unit 904 is configured to perform fusion of the propagation loss and the interaction loss negatively correlated to the first similarity, to obtain a model loss of the recommended model; and updating the model parameters in the recommendation model and the first characteristics of the plurality of objects based on the model loss to obtain the trained recommendation model.

In some embodiments, the apparatus further comprises:

a fourth determining unit configured to perform determining a third object corresponding to each first object from second behavior information, the second behavior information including new-added behavior data and a plurality of pieces of historical behavior data, the third object being another object in the second behavior information except the first object; determining a negative sampling loss negatively correlated with a third similarity, the third similarity being determined based on a similarity between the interaction feature of the first object and a context feature of a third object, the context feature characterizing features of the third object under the influence of other objects;

a training unit 904 configured to perform training of the recommendation model based on the propagation loss and the negative sampling loss negatively correlated to the first similarity.

In some embodiments, the apparatus further comprises:

a fourth determining unit configured to perform determining a third object corresponding to each first object from second behavior information, the second behavior information including new-added behavior data and a plurality of pieces of historical behavior data, the third object being another object in the second behavior information except the first object;

determining a negative sampling loss negatively correlated with a third similarity, the third similarity being determined based on a similarity between the interaction feature of the first object and a context feature of a third object, the context feature characterizing features of the third object under the influence of other objects;

a training unit 904 configured to perform training of the recommendation model based on the propagation loss, the interaction loss and the negative sampling loss negatively correlated to the first similarity.

In some embodiments, the second obtaining unit 901 is configured to perform sampling on second behavior information according to two first objects in the new added behavior data to obtain the first behavior information, where the second behavior information includes the new added behavior data and a plurality of pieces of historical behavior data.

In some embodiments, the second obtaining unit 901 is configured to perform obtaining a sampling mode set, where the sampling mode set includes a plurality of sampling modes; determining a sampling mode of each first object from the sampling mode set; for each first object, sampling is carried out in the second behavior information according to the determined sampling mode from the first object, and historical behavior data and newly-added behavior data obtained by sampling form the first behavior information.

In some embodiments, the first feature of the second object comprises a contextual feature, the contextual feature characterizing features of the second object under the influence of other objects; the device still includes:

a fifth determining unit configured to perform determining a similarity between the influence feature and the context feature of each second object; and fusing the similarity of at least one second object to obtain a first similarity.

In some embodiments, the apparatus further comprises:

a training unit 904 further configured to perform obtaining a plurality of pieces of sample behavior data, each piece of sample behavior data being used to represent an interaction behavior between two sample objects belonging to different types; dividing a plurality of sample behavior data into a plurality of sample sets according to the sequence of the occurrence time from far to near, wherein the number of the sample behavior data contained in each sample set is the same; and training the recommendation model in sequence based on a plurality of sample sets.

With regard to the apparatus in the above-described embodiment, the specific manner in which each unit performs the operation has been described in detail in the embodiment related to the method, and will not be described in detail here.

In an exemplary embodiment, an electronic device is provided that includes one or more processors, and a memory to store instructions executable by the one or more processors; wherein the one or more processors are configured to perform the recommendation model processing method or the media content recommendation method in the above embodiments.

In some embodiments, the electronic device is provided as a terminal. Fig. 10 is a block diagram illustrating a structure of a terminal 1000 according to an exemplary embodiment. The terminal 1000 can be a portable mobile terminal such as: a smart phone, a tablet computer, an MP3 player (Moving Picture Experts Group Audio Layer III, motion video Experts compression standard Audio Layer 3), an MP4 player (Moving Picture Experts Group Audio Layer IV, motion video Experts compression standard Audio Layer 4), a notebook computer, or a desktop computer. Terminal 1000 can also be referred to as user equipment, portable terminal, laptop terminal, desktop terminal, or the like by other names.

Terminal 1000 can include: a processor 1001 and a memory 1002.

Processor 1001 may include one or more processing cores, such as a 4-core processor, an 8-core processor, and so forth. The processor 1001 may be implemented in at least one hardware form of a DSP (Digital Signal Processing), an FPGA (Field-Programmable Gate Array), and a PLA (Programmable Logic Array). The processor 1001 may also include a main processor and a coprocessor, where the main processor is a processor for Processing data in an awake state, and is also referred to as a Central Processing Unit (CPU); a coprocessor is a low power processor for processing data in a standby state. In some embodiments, the processor 1001 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content that the display screen needs to display. In some embodiments, the processor 1001 may further include an AI (Artificial Intelligence) processor for processing a computing operation related to machine learning.

Memory 1002 may include one or more computer-readable storage media, which may be non-transitory. The memory 1002 may also include high-speed random access memory, as well as non-volatile memory, such as one or more magnetic disk storage devices, flash memory storage devices. In some embodiments, a non-transitory computer readable storage medium in memory 1002 is used to store at least one executable instruction for execution by processor 1001 to implement a recommendation model processing method or a media content recommendation method provided by method embodiments in the present disclosure.

In some embodiments, terminal 1000 can also optionally include: a peripheral interface 1003 and at least one peripheral. The processor 1001, memory 1002 and peripheral interface 1003 may be connected by a bus or signal line. Various peripheral devices may be connected to peripheral interface 1003 via a bus, signal line, or circuit board. Specifically, the peripheral device includes: at least one of radio frequency circuitry 1004, display screen 1005, camera assembly 1006, audio circuitry 1007, positioning assembly 1008, and power supply 1009.

The peripheral interface 1003 may be used to connect at least one peripheral related to I/O (Input/Output) to the processor 1001 and the memory 1002. In some embodiments, processor 1001, memory 1002, and peripheral interface 1003 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 1001, the memory 1002, and the peripheral interface 1003 may be implemented on separate chips or circuit boards, which are not limited by this embodiment.

The Radio Frequency circuit 1004 is used for receiving and transmitting RF (Radio Frequency) signals, also called electromagnetic signals. The radio frequency circuitry 1004 communicates with communication networks and other communication devices via electromagnetic signals. The radio frequency circuit 1004 converts an electrical signal into an electromagnetic signal to transmit, or converts a received electromagnetic signal into an electrical signal. Optionally, the radio frequency circuit 1004 comprises: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a subscriber identity module card, and so forth. The radio frequency circuit 1004 may communicate with other terminals via at least one wireless communication protocol. The wireless communication protocols include, but are not limited to: the world wide web, metropolitan area networks, intranets, generations of mobile communication networks (2G, 3G, 4G, and 5G), Wireless local area networks, and/or WiFi (Wireless Fidelity) networks. In some embodiments, the radio frequency circuit 1004 may also include NFC (Near Field Communication) related circuits, which are not limited by this disclosure.

The display screen 1005 is used to display a UI (User Interface). The UI may include graphics, text, icons, video, and any combination thereof. When the display screen 1005 is a touch display screen, the display screen 1005 also has the ability to capture touch signals on or over the surface of the display screen 1005. The touch signal may be input to the processor 1001 as a control signal for processing. At this point, the display screen 1005 may also be used to provide virtual buttons and/or a virtual keyboard, also referred to as soft buttons and/or a soft keyboard. In some embodiments, display screen 1005 can be one, disposed on a front panel of terminal 1000; in other embodiments, display 1005 can be at least two, respectively disposed on different surfaces of terminal 1000 or in a folded design; in other embodiments, display 1005 can be a flexible display disposed on a curved surface or a folded surface of terminal 1000. Even more, the display screen 1005 may be arranged in a non-rectangular irregular figure, i.e., a shaped screen. The Display screen 1005 may be made of LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), and the like.

The camera assembly 1006 is used to capture images or video. Optionally, the camera assembly 1006 includes a front camera and a rear camera. The front camera is arranged on the front panel of the terminal, and the rear camera is arranged on the back of the terminal. In some embodiments, the number of the rear cameras is at least two, and each rear camera is any one of a main camera, a depth-of-field camera, a wide-angle camera and a telephoto camera, so that the main camera and the depth-of-field camera are fused to realize a background blurring function, and the main camera and the wide-angle camera are fused to realize panoramic shooting and VR (Virtual Reality) shooting functions or other fusion shooting functions. In some embodiments, camera assembly 1006 may also include a flash. The flash lamp can be a monochrome temperature flash lamp or a bicolor temperature flash lamp. The double-color-temperature flash lamp is a combination of a warm-light flash lamp and a cold-light flash lamp, and can be used for light compensation at different color temperatures.

The audio circuit 1007 may include a microphone and a speaker. The microphone is used for collecting sound waves of a user and the environment, converting the sound waves into electric signals, and inputting the electric signals to the processor 1001 for processing or inputting the electric signals to the radio frequency circuit 1004 for realizing voice communication. For stereo sound collection or noise reduction purposes, multiple microphones can be provided, each at a different location of terminal 1000. The microphone may also be an array microphone or an omni-directional pick-up microphone. The speaker is used to convert electrical signals from the processor 1001 or the radio frequency circuit 1004 into sound waves. The loudspeaker can be a traditional film loudspeaker or a piezoelectric ceramic loudspeaker. When the speaker is a piezoelectric ceramic speaker, the speaker can be used for purposes such as converting an electric signal into a sound wave audible to a human being, or converting an electric signal into a sound wave inaudible to a human being to measure a distance. In some embodiments, the audio circuit 1007 may also include a headphone jack.

A Location component 1008 is employed to locate a current geographic Location of terminal 1000 for purposes of navigation or LBS (Location Based Service). The Positioning component 1008 may be a Positioning component based on a Global Positioning System (GPS) in the united states, a beidou System in china, a greiner Positioning System in russia, or a galileo Positioning System in the european union.

Power supply 1009 is used to supply power to various components in terminal 1000. The power source 1009 may be alternating current, direct current, disposable batteries, or rechargeable batteries. When the power source 1009 includes a rechargeable battery, the rechargeable battery may be a wired rechargeable battery or a wireless rechargeable battery. The wired rechargeable battery is a battery charged through a wired line, and the wireless rechargeable battery is a battery charged through a wireless coil. The rechargeable battery may also be used to support fast charge technology.

In some embodiments, terminal 1000 can also include one or more sensors 1010. The one or more sensors 1010 include, but are not limited to: acceleration sensor 1011, gyro sensor 1012, pressure sensor 1013, fingerprint sensor 1014, optical sensor 1015, and proximity sensor 1016.

Acceleration sensor 1011 can detect acceleration magnitudes on three coordinate axes of a coordinate system established with terminal 1000. For example, the acceleration sensor 1011 may be used to detect components of the gravitational acceleration in three coordinate axes. The processor 1001 may control the display screen 1005 to display the user interface in a landscape view or a portrait view according to the gravitational acceleration signal collected by the acceleration sensor 1011. The acceleration sensor 1011 may also be used for acquisition of motion data of a game or a user.

The gyro sensor 1012 may detect a body direction and a rotation angle of the terminal 1000, and the gyro sensor 1012 and the acceleration sensor 1011 may cooperate to acquire a 3D motion of the user on the terminal 1000. From the data collected by the gyro sensor 1012, the processor 1001 may implement the following functions: motion sensing (such as changing the UI according to a user's tilting operation), image stabilization at the time of photographing, game control, and inertial navigation.

Pressure sensor 1013 can be disposed on a side frame of terminal 1000 and/or underneath display screen 1005. When pressure sensor 1013 is disposed on a side frame of terminal 1000, a user's grip signal on terminal 1000 can be detected, and processor 1001 performs left-right hand recognition or shortcut operation according to the grip signal collected by pressure sensor 1013. When the pressure sensor 1013 is disposed at a lower layer of the display screen 1005, the processor 1001 controls the operability control on the UI interface according to the pressure operation of the user on the display screen 1005. The operability control comprises at least one of a button control, a scroll bar control, an icon control and a menu control.

The fingerprint sensor 1014 is used to collect a fingerprint of the user, and the processor 1001 identifies the user according to the fingerprint collected by the fingerprint sensor 1014, or the fingerprint sensor 1014 identifies the user according to the collected fingerprint. Upon identifying that the user's identity is a trusted identity, the processor 1001 authorizes the user to perform relevant sensitive operations including unlocking a screen, viewing encrypted information, downloading software, paying, and changing settings, etc. Fingerprint sensor 1014 may be disposed on a front, back, or side of terminal 1000. When a physical key or vendor Logo is provided on terminal 1000, fingerprint sensor 1014 can be integrated with the physical key or vendor Logo.

The optical sensor 1015 is used to collect the ambient light intensity. In one embodiment, the processor 1001 may control the display brightness of the display screen 1005 according to the ambient light intensity collected by the optical sensor 1015. Specifically, when the ambient light intensity is high, the display brightness of the display screen 1005 is increased; when the ambient light intensity is low, the display brightness of the display screen 1005 is turned down. In another embodiment, the processor 1001 may also dynamically adjust the shooting parameters of the camera assembly 1006 according to the intensity of the ambient light collected by the optical sensor 1015.

Proximity sensor 1016, also known as a distance sensor, is disposed on a front panel of terminal 1000. Proximity sensor 1016 is used to gather the distance between the user and the front face of terminal 1000. In one embodiment, when proximity sensor 1016 detects that the distance between the user and the front surface of terminal 1000 is gradually reduced, processor 1001 controls display screen 1005 to switch from a bright screen state to a dark screen state; when proximity sensor 1016 detects that the distance between the user and the front of terminal 1000 is gradually increased, display screen 1005 is controlled by processor 1001 to switch from a breath-screen state to a bright-screen state.

Those skilled in the art will appreciate that the configuration shown in FIG. 10 is not intended to be limiting and that terminal 1000 can include more or fewer components than shown, or some components can be combined, or a different arrangement of components can be employed.

In other embodiments, the electronic device is provided as a server. Fig. 11 is a block diagram illustrating a server 1100, which may have a relatively large difference due to different configurations or performances, according to an exemplary embodiment, and may include one or more processors (CPUs) 1101 and one or more memories 1102, where the memory 1102 stores at least one executable instruction, and the at least one executable instruction is loaded and executed by the processors 1101 to implement the methods provided by the above-mentioned method embodiments. Of course, the server may also have components such as a wired or wireless network interface, a keyboard, and an input/output interface, so as to perform input/output, and the server may also include other components for implementing the functions of the device, which are not described herein again.

In an exemplary embodiment, there is also provided a computer-readable storage medium having instructions stored thereon, which, when executed by a processor of an electronic device, enable the electronic device to perform the recommendation model processing method or the media content recommendation method described above. Alternatively, the computer-readable storage medium may be a ROM (Read Only Memory), a RAM (Random Access Memory), a CD-ROM (Compact Disc Read-Only Memory), a magnetic tape, a floppy disk, an optical data storage device, and the like.

In an exemplary embodiment, a computer program product is also provided, which includes a computer program executed by a processor to implement the recommendation model processing method or the media content recommendation method described above.

In some embodiments, the computer program according to the embodiments of the present application may be deployed to be executed on one computer device or on multiple computer devices located at one site, or may be executed on multiple computer devices distributed at multiple sites and interconnected by a communication network, and the multiple computer devices distributed at the multiple sites and interconnected by the communication network may constitute a block chain system.

Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure herein. This disclosure is intended to cover any variations, uses, or adaptations of the disclosure following, in general, the principles of the disclosure and including such departures from the present disclosure as come within known or customary practice within the art to which the disclosure pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.

It will be understood that the present disclosure is not limited to the precise arrangements described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method for recommending media contents, the method comprising:

2. The media content recommendation method according to claim 1, wherein the account characteristics include a first memory characteristic and a second memory characteristic of the first user account, the first memory characteristic characterizes a long-term characteristic of the first user account, and the second memory characteristic characterizes a short-term characteristic of the first user account; the content features comprise first memory features and second memory features of the first media content, the first memory features of the first media content characterize the first media content in a long term, the second memory features of the first media content characterize the first media content in a short term;

3. A recommendation model processing method, the method comprising:

4. The recommendation model processing method according to claim 3, wherein said determining, for each of said first objects, an influence feature of said second object based on a first feature of said first object and an interaction time difference of said second object related to said first object comprises:

5. The recommendation model processing method according to claim 4, wherein determining the impact characteristics of the second object related to the first object based on the interaction characteristics of the first object and the interaction time difference of the second object belonging to the same historical behavior data as the first object comprises:

6. An apparatus for recommending media contents, said apparatus comprising:

7. A recommendation model processing apparatus, characterized in that the apparatus comprises:

8. An electronic device, characterized in that the electronic device comprises:

one or more processors;

a memory for storing the one or more processor-executable instructions;

wherein the one or more processors are configured to perform the media content recommendation method of any of claims 1-2 or the recommendation model processing method of any of claims 3-5.

9. A computer-readable storage medium, wherein instructions in the computer-readable storage medium, when executed by a processor of an electronic device, enable the electronic device to perform the media content recommendation method of any one of claims 1-2, or the recommendation model processing method of any one of claims 3-5.

10. A computer program product comprising a computer program, characterized in that the computer program, when executed by a processor, implements a media content recommendation method according to any one of claims 1 to 2, or is capable of performing a recommendation model processing method according to any one of claims 3 to 5.