CN117251622A

CN117251622A - Method, device, computer equipment and storage medium for recommending objects

Info

Publication number: CN117251622A
Application number: CN202310968728.2A
Authority: CN
Inventors: 赵鑫; 谢若冰; 张君杰; 孙文奇; 周杰
Original assignee: Tencent Technology Shenzhen Co Ltd; Renmin University of China
Current assignee: Tencent Technology Shenzhen Co Ltd; Renmin University of China
Priority date: 2023-08-02
Filing date: 2023-08-02
Publication date: 2023-12-19

Abstract

The present application relates to a method, apparatus, computer device, storage medium and computer program product for object recommendation. The method comprises the following steps: acquiring interactive objects triggered by a target user to perform interactive operation in a historical time period, and determining evaluation information fed back by the target user for each interactive object respectively; extracting object preference characteristics of the target user for each interactive object through each piece of evaluation information, wherein the object preference characteristics are used for representing the preference of the target user for the interactive object; generating object preference distribution characteristics of the target user based on the object preference characteristics of the target user for each interactive object, wherein the object preference distribution characteristics are used for representing preference object characteristics of objects preferred by the target user; a target object is determined from the candidate objects based on the object preference distribution characteristics, and the target object is recommended to a target user. By adopting the method, the accuracy of object recommendation can be improved.

Description

Method, device, computer equipment and storage medium for recommending objects

Technical Field

The present application relates to information processing technology, and in particular, to a method, an apparatus, a computer device, and a storage medium for object recommendation.

Background

With the rapid development of computing technology and internet technology, people tend to acquire network information through the internet, and many internet platforms often recommend objects to users in order to better serve and attract users, so how to more accurately recommend objects of interest to users becomes an important means for improving the experience of users browsing internet platforms.

Currently, the user is generally predicted based on his or her historical interaction behavior to predict the objects that the user prefers, and then the preferred objects are recommended to the user. However, the user preferences predicted in the foregoing manner can only be made from the dimensions of the historical interaction behavior, and thus the resulting predicted results may not conform to the user's actual preferences for the object, thereby reducing the accuracy of the object recommendation. Therefore, how to improve the accuracy of object recommendation is a problem to be solved.

Disclosure of Invention

In view of the foregoing, it is desirable to provide an object recommendation method, apparatus, computer device, and storage medium that can improve accuracy of object recommendation.

In a first aspect, the present application provides a method of object recommendation. The method comprises the following steps:

Acquiring interactive objects triggered by a target user to perform interactive operation in a historical time period, and determining evaluation information fed back by the target user for each interactive object respectively;

extracting object preference characteristics of the target user for each interactive object through each piece of evaluation information, wherein the object preference characteristics are used for representing the preference of the target user for the interactive object;

generating object preference distribution characteristics of the target user based on the object preference characteristics of the target user for each interactive object, wherein the object preference distribution characteristics are used for representing preference object characteristics of objects preferred by the target user;

a target object is determined from the candidate objects based on the object preference distribution characteristics, and the target object is recommended to a target user.

In a second aspect, the application further provides an object recommendation device. The device comprises:

the evaluation information determining module is used for acquiring interactive objects triggered by the target user to perform interactive operation in a historical time period and determining evaluation information fed back by the target user for each interactive object respectively;

the object preference feature extraction module is used for extracting object preference features of the target user aiming at each interactive object through each piece of evaluation information, wherein the object preference features are used for representing the preference of the target user on the interactive object;

The object preference distribution feature generation module is used for generating object preference distribution features of the target user based on the object preference features of the target user for each interactive object, wherein the object preference distribution features are used for representing preference object features of objects preferred by the target user;

and the object recommendation module is used for determining a target object from the candidate objects based on the object preference distribution characteristics and recommending the target object to a target user.

In a third aspect, the present application also provides a computer device. The computer device comprises a memory storing a computer program and a processor which when executing the computer program performs the steps of:

In a fourth aspect, the present application also provides a computer-readable storage medium. The computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of:

In a fifth aspect, the present application also provides a computer program product. The computer program product comprises a computer program which, when executed by a processor, implements the steps of:

The object recommendation method, the device, the computer equipment, the storage medium and the computer program product are characterized in that firstly, interactive objects which are triggered by a target user to perform interactive operation in a historical time period are obtained, evaluation information fed back by the target user for each interactive object is determined, through each evaluation information, object preference characteristics of the target user for each interactive object are extracted, the object preference characteristics are used for representing preferences of the target user for each interactive object, then, object preference distribution characteristics of the target user are generated based on the object preference characteristics of the target user for each interactive object, the object preference distribution characteristics are used for representing preference object characteristics of the objects preferred by the target user, finally, the target object is determined from each candidate object based on the object preference distribution characteristics, and the target object is recommended to the target user. According to the object recommendation method, the evaluation information fed back by the target user on the interactive object in the history time period is usually freely written to the interactive object by the target user, so that personalized preferences of the target user can be obtained more directly and accurately through the evaluation information, and then the object preference distribution characteristics are generated by extracting the object preference characteristics directly and accurately through the evaluation information.

Drawings

FIG. 1 is an application environment diagram of a method of object recommendation in one embodiment;

FIG. 2 is a flow diagram of a method of object recommendation in one embodiment;

FIG. 3 is a partial flow diagram of generating object preference distribution features in one embodiment;

FIG. 4 is a flow diagram of determining interactive behavior features in one embodiment;

FIG. 5 is a partial flow diagram of generating object preference distribution features in another embodiment;

FIG. 6 is a flow diagram of a method for generating object preference distribution features from object preference features, interaction behavior features, and object features in one embodiment;

FIG. 7 is a complete flow diagram of generating object preference distribution features in one embodiment;

FIG. 8 is a flow diagram of extracting object preference features and object features in one embodiment;

FIG. 9 is a flow diagram of determining a target object in one embodiment;

FIG. 10 is a flow diagram of a method for obtaining a preference prediction model in one embodiment;

FIG. 11 is a flow diagram of adjusting model parameters of a preference prediction model in one embodiment;

FIG. 12 is a flowchart of a method for obtaining a preference prediction model according to another embodiment;

FIG. 13 is a complete flow diagram of adjusting model parameters of a preference prediction model in one embodiment;

FIG. 14 is a flow diagram of a process for obtaining a negative sample of an interactive object in one embodiment;

FIG. 15 is a complete flow diagram of a method of object recommendation in one embodiment;

FIG. 16 is a block diagram of an object recommendation device in one embodiment;

FIG. 17 is a block diagram showing an object recommendation apparatus according to another embodiment;

FIG. 18 is a block diagram showing an object recommending apparatus according to still another embodiment;

FIG. 19 is a block diagram showing an object recommending apparatus according to still another embodiment;

fig. 20 is an internal structural view of a computer device in one embodiment.

Detailed Description

In order to make the objects, technical solutions and advantages of the present application more apparent, the present application will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the present application.

With the rapid development of computing technology and internet technology, people tend to acquire network information through the internet, and many internet platforms often recommend objects to users in order to better serve and attract users, so how to more accurately recommend objects of interest to users becomes an important means for improving the experience of users browsing internet platforms. Currently, a user prefers an object based on a historical interaction behavior of the user is usually predicted, then the preferred object is recommended to the user, namely, the historical implicit preference of the user is specifically extracted from the historical interaction behavior, so that the user prefers the object to be recommended based on the matching of the historical implicit preference and the object to be recommended, however, the current intention and preference of the user cannot be reliably reflected from the historical interaction behavior, namely, the historical explicit preference cannot be accurately and reliably extracted from the historical interaction behavior, namely, the interactive object of the user triggering the interaction operation is not necessarily the object preferred by the user, so that the predicted user preference can only be carried out from the dimension of the historical interaction behavior, and thus the obtained prediction result may not conform to the actual preference of the user for the object, thereby reducing the accuracy of object recommendation.

Therefore, the embodiment of the application provides an object recommendation method capable of improving accuracy of object recommendation. The method for recommending the object, provided by the embodiment of the application, can be applied to an application environment shown in fig. 1. Wherein the terminal 102 communicates with the server 104 via a network. The data storage system may store data that the server 104 needs to process. The data storage system may be integrated on the server 104 or may be located on the cloud or other servers.

Specifically, taking the application to the server 104 as an example, the server 104 obtains the interactive objects that are triggered by the target user to perform interactive operation in a historical time period, determines evaluation information that is fed back by the target user for each interactive object, then, the server 104 extracts object preference characteristics of the target user for each interactive object through each evaluation information, the object preference characteristics are used for representing preferences of the target user for each interactive object, and generates object preference distribution characteristics of the target user for representing preference object characteristics of the target user for each interactive object based on the object preference characteristics of the target user, finally, determines the target object from each candidate object based on the object preference distribution characteristics, and recommends the target object to the target user, and the server 104 can send the target object to the terminal 102 through communication connection between the terminal 102 and the terminal 102 used by the target user, so that the terminal 102 recommends the target object.

The personalized preference of the target user can be obtained more directly and accurately through the evaluation information, and then the object preference characteristics are extracted through the direct and accurate evaluation information, so that the object preference distribution characteristics are generated, and when the object recommendation is performed to the target user, the object preference distribution characteristics, namely the object characteristics corresponding to the object with the preference of the target user, are considered, and the target object for recommendation is selected from the candidate objects, so that the accuracy of the object recommendation is improved.

The terminal 102 may be, but not limited to, various desktop computers, notebook computers, smart phones, tablet computers, internet of things devices, and portable wearable devices, where the internet of things devices may be smart speakers, smart televisions, smart air conditioners, smart vehicle devices, and the like. The portable wearable device may be a smart watch, smart bracelet, headset, or the like. The server 104 may be implemented as a stand-alone server or as a server cluster of multiple servers. And the method for recommending the object provided by the application embodiment can be applied to various scenes, including but not limited to cloud technology, artificial intelligence and the like.

Further, since the method for object recommendation provided in the embodiments of the present application further relates to artificial intelligence (Artificial Intelligence, AI) technology, the AI technology will be briefly described below: AI is a theory, method, technique, and application system that utilizes a digital computer or a digital computer-controlled machine to simulate, extend, and extend human intelligence, sense the environment, acquire knowledge, and use knowledge to obtain optimal results. In other words, artificial intelligence is an integrated technology of computer science that attempts to understand the essence of intelligence and to produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence, i.e. research on design principles and implementation methods of various intelligent machines, enables the machines to have functions of sensing, reasoning and decision.

The artificial intelligence technology is a comprehensive subject, and relates to the technology with wide fields, namely the technology with a hardware level and the technology with a software level. Artificial intelligence infrastructure technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and other directions.

Machine Learning (ML) is a multi-domain interdisciplinary, involving multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It is specially studied how a computer simulates or implements learning behavior of a human to acquire new knowledge or skills, and reorganizes existing knowledge structures to continuously improve own performance. Machine learning is the core of artificial intelligence, a fundamental approach to letting computers have intelligence, which is applied throughout various areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, confidence networks, reinforcement learning, transfer learning, induction learning, teaching learning, and the like.

Further, taking machine learning as an example, a first interaction object sample triggered by a target user to perform interaction operation in a first historical time period is obtained, a second interaction object sample triggered by the target user to perform interaction operation in a second historical time period is obtained, the second historical time period is the next historical time period adjacent to the first historical time period, then a first evaluation information sample of the target user for each first interaction object sample is determined, so that a first object preference feature of the target user for each first interaction object sample is extracted through an initial preference prediction model based on each first evaluation information sample, object attribute features corresponding to each second interaction object sample are extracted, a first object preference distribution feature of the target user is generated through the initial preference prediction model based on each first object preference feature, object attribute features corresponding to each second interaction object sample are extracted, and finally model parameters of the initial preference prediction model are adjusted through the first object preference distribution feature and each object attribute feature, so that a preference prediction model is obtained. Thus, in the object recommendation process, object preference features can be extracted and object preference distribution features generated by the preference prediction model. And in practical applications, the scenario of making object recommendation may include, but is not limited to: commodity recommendation, text recommendation, video recommendation, audio recommendation, virtual data recommendation, and the like, are not particularly limited herein.

The scheme provided by the embodiment of the application relates to artificial neural network and other technologies of artificial intelligence, and is specifically described through the following embodiments: in one embodiment, as shown in fig. 2, a method for object recommendation is provided, and is illustrated by taking the method applied to the server 104 in fig. 1 as an example, it is understood that the method may also be applied to the terminal 102, and may also be applied to a system including the terminal 102 and the server 104, and implemented through interaction between the terminal 102 and the server 104. In this embodiment, the method includes the steps of:

step 202, acquiring interactive objects triggered by a target user to perform interactive operation in a historical time period, and determining evaluation information fed back by the target user for each interactive object.

The interactive objects can be one or more, and the interactive objects can be commodities, texts, audios and videos, virtual data and the like. It can be seen that the interactive object has a corresponding relationship with the operation type of the triggered interactive operation, and the operation type of the interactive operation is determined based on the application requirement. For example: in the case of application to a commodity recommendation scenario, then the interactive operation triggered by the target user may be a commodity settlement, and the interactive object may be a commodity that has been subjected to settlement. Alternatively, where applied to a video recommendation scenario, then the triggered interaction may be to view video or the like, while the interaction object may have been viewed. Therefore, the specific interactive object and the triggered specific interactive operation are not limited in this embodiment.

Next, the evaluation information includes at least: the personalized preference of the target user is specific to the description information and noise information of each interactive object, the personalized preference of the target user is specific to the information of object preference dimension, and the object preference dimension is specific to the non-visual description expressed by the target user to the object, and is mainly used for expressing emotion and tendency of the object. The description information is specifically information of an object description dimension, and the object description dimension is specifically visual description expressed by a target user to an object, namely, the object description dimension does not carry emotion and tendency of the object, but describes the essence of the object, so that the object description dimension can be directly the object information for describing the object. The noise information is information independent of the personalized preferences and the descriptive information. And the evaluation information may include data information in text format, may also include data information in picture format, or even data information in audio format. It is to be understood that in practical application, the situation that the evaluation information of the target user for a certain interaction object is empty may also occur, which is not limited herein.

Specifically, in a scenario in which object recommendation is performed for a target user, a server can acquire an interactive object in which an interactive operation is triggered by the target user in a history period. That is, the server may obtain, from the data information stored in the terminal, an interactive object that is triggered by the target user to perform an interactive operation during the history period through a communication connection with the terminal used by the target user. Or after each user triggers the interactive operation on each object in the historical time period, the interactive object of each user can be transmitted to the server in real time, and the server stores the interactive object of each user in the data storage system, so that when object recommendation is required for the target user, the interactive object of each user, which is triggered by the target user in the historical time period, is determined from the interactive objects of each user stored in the data storage system. For example, if the target user triggers an interactive operation on the object A1, the object A2, and the object A3, respectively, in the history period, the object A1, the object A2, and the object A3 may be determined as interactive objects, respectively. And the manner of determining the interactive object is not particularly limited in this embodiment.

Further, the server further acquires evaluation information fed back by the target user for each interactive object. That is, when the target user feeds back for the interactive object, the server may acquire evaluation information including various information. For example, taking the application to a commodity recommendation scenario as an example, the user settles and purchases commodity 1, commodity 2 and commodity 3 in a historical time period, the user can feedback the evaluation to commodity 1, and then the information included in the feedback evaluation is evaluation information, for example, the evaluation information fed back to commodity 1 is specifically "commodity 1 size standard", but the price is somewhat expensive, and the price is hoped to be cheaper. And the user may feed back the evaluation for the commodity 2, such as the evaluation information fed back for the commodity 2 is specifically "commodity 2 is very good in light blue but of poor quality". And the user can also feed back the evaluation to the commodity 3, for example, the evaluation information fed back to the commodity 3 is specifically "the commodity 3 is not right in size and is difficult to wear.

Secondly, the data information included in the evaluation information may be null, and the target user may not perform feedback evaluation on the interaction object, and at this time, the evaluation information acquired by the server is null. Taking the commodity recommendation scene as an example again, if the user does not feed back any information to the commodity 1 after the user settles and purchases the commodity 1, the server can still acquire the evaluation information fed back by the user to the commodity 1, but the data information in the evaluation information fed back by the user to the commodity 1 is empty.

It will be appreciated that, in the present embodiment, the one interaction described in the present embodiment may be used to characterize an operation of obtaining an execution result, and then the one interaction may be specifically an instant interaction of obtaining an execution result, for example, an operation of settling a commodity. Secondly, the one-time interaction described in the embodiment may also be a continuous interaction performed during the execution time interval, for example, an operation of playing a single video, and then, during the video duration of the single video, the continuous playing of the video is also an interaction operation, or an operation of viewing a single text, and then, the continuous viewing of the single text by the target user is also an interaction operation. Therefore, the present embodiment does not specifically limit the interactive operation.

Step 204, extracting object preference characteristics of the target user for each interactive object through each piece of evaluation information, wherein the object preference characteristics are used for representing the preference of the target user for the interactive object.

Wherein the object preference feature is used to characterize the preferences of the target user for the interactive object. Preferences are non-intuitive descriptions, so the preferences are emotions and trends of the target user on the interactive objects, and the preferences can be specifically expressed as the preference degree of the target user on the interactive objects. Thus, the preference may be described by a percentage, such as: the closer the preference is to 100%, the higher the preference of the target user for the interactive object is, the more satisfied and liked by the target user for the interactive object is, and the closer the preference is to 0%, the lower the preference of the target user for the interactive object is, the less satisfied or disliked by the target user for the interactive object is. Second, preferences may also be described in terms of scoring under a percentile, e.g., closer to 100 the preferences indicate a higher preference of the target user for the interactive object, and closer to 0 the preferences indicate a lower preference of the target user for the interactive object.

Specifically, the server extracts object preference characteristics of the target user for each interactive object through each piece of evaluation information. Specifically, general semantic features corresponding to the evaluation information can be extracted first, and then object preference features can be extracted from the general semantic features corresponding to the evaluation information. As can be seen from the foregoing description, since the evaluation information includes at least: the personalized preference of the target user, the description information of each interactive object and the noise information, so that the server can determine the general semantic features corresponding to the personalized preference of the target user from the general semantic features corresponding to the evaluation information, and further determine the object preference features.

It may be appreciated that, if the evaluation information of the target user for a certain interactive object is null, the object preference feature may be obtained based on the interactive behavior information generated when the target user triggers the interactive operation for the interactive object.

In step 206, based on the object preference characteristics of the target user for each interactive object, object preference distribution characteristics of the target user are generated, the object preference distribution characteristics being used to characterize preferred object characteristics of the objects preferred by the target user.

Wherein the object preference distribution feature is used to characterize preferred object features of objects preferred by the target user, and the preferred object features may include at least object types of the objects. Specifically, the server encodes the attention of the object preference characteristics of the evaluation information fed back by the target user based on the object preference characteristics of the target user for each interactive object, so that the characteristic representation distance between the object preference characteristics can be obtained, and the object preference distribution characteristics capable of reflecting the user preference can be generated. For example, taking the example of the application to the commodity recommendation scenario described in the foregoing embodiment again, the commodities 1, 2 and 3 are all shoes, and the obtained object preference distribution characteristics may represent that the object preferred by the target user is "shoes", "comfortable", "good quality" and "low price" based on the foregoing evaluation information fed back for the commodities 1, 2 and 3.

In step 208, a target object is determined from the candidate objects based on the object preference distribution characteristics, and the target object is recommended to the target user.

The candidate object is an object belonging to a candidate object set, and the candidate object is an object preset based on an application scene, for example, in the case of being applied to a commodity recommendation scene, the candidate object set can be a set formed by all commodities under commodity shopping software, and then the candidate object is any commodity under commodity shopping software. Second, as in the case of application to the paper text scene, the candidate set may be a set of all paper texts included in the paper text website, and then the candidate is any paper text under the paper text website.

Specifically, the server extracts candidate object features of each candidate object, then determines a target object through the similarity between each candidate object feature and the object preference distribution feature, and recommends the determined target object to the target user. In this embodiment, the server 104 in fig. 1 is taken as an example, so that the server may send the determined target object to the terminal through a communication connection with the terminal used by the target user, so that the terminal used by the target user recommends the target object.

It is to be understood that the foregoing examples have been provided merely for the understanding of the present invention and are not to be construed as limiting the invention in any way.

In the method for recommending the object, the evaluation information fed back by the target user to the interactive object in the history period is usually freely written to the interactive object by the target user, so that the personalized preference of the target user can be obtained more directly and accurately through the evaluation information, and then the object preference distribution characteristics are generated by extracting the object preference characteristics directly and accurately through the evaluation information.

In one embodiment, as shown in fig. 3, the method for recommending objects further includes:

step 302, determining interaction behavior information generated each time the interaction operation is triggered, and determining interaction behavior characteristics based on each interaction behavior information.

Wherein, the interaction behavior information at least comprises: the object type of the interactive object, the object information of the interactive object, etc. of the interactive object, which is triggered by the interactive operation. For example, the object type of the interactive object that is triggered to the interactive operation is sports shoes, and the object information of the interactive object may be a sports shoe brand, a sports shoe size, a sports shoe color, or the like. Or, the object type of the interactive object triggered by the interactive operation is a paper, and the object information of the interactive object may be a paper name, a paper author, a paper field, and the like. And secondly, the interactive behavior features are at least used for representing the behavior information features of each interactive behavior information and the associated information features among the interactive behavior information.

Specifically, after the server acquires the interactive object whose interaction operation is triggered by the target user in the history period, the server can further determine the interaction behavior information generated each time the interaction operation is triggered. If the server obtains the interactive object through communication connection with the terminal used by the target user, the server can further obtain the interactive behavior information generated each time the interactive operation is triggered from the terminal through communication with the terminal used by the target user. Alternatively, if the server obtains the interactive object from the data storage system, the server may also obtain the interaction behavior information generated each time the interaction is triggered from the data storage system. Therefore, the present embodiment does not specifically limit the manner of determining the interaction behavior information.

Further, the server determines the interaction behavior characteristics based on the interaction behavior information, i.e. the server codes the attention of the interaction behavior information to obtain the interaction behavior characteristics. The process of attention encoding by the server at this time is implemented based on a preference prediction model.

Based on the above, generating the object preference distribution feature of the target user based on the object preference feature of the target user for each interactive object, including:

Step 304, generating object preference distribution characteristics based on the object preference characteristics of the target user for each interactive object and the interactive behavior characteristics.

Specifically, the server generates object preference distribution characteristics based on object preference characteristics of the target user for each interactive object, and interactive behavior characteristics. Namely, the server codes the attention of each object preference feature and each interaction behavior feature, so that not only the feature representation distance between the object preference features can be obtained, but also the behavior information features of each interaction behavior information and the associated information features between the interaction behavior information can be considered to generate object preference distribution features, and at the moment, the object preference distribution features can be specifically characterized: preference object characteristics of objects preferred by the target user over a historical period of time.

In this embodiment, since the object preference feature of the target user for each interactive object can reflect the intrinsic preference of the target user, and the encoding is performed on the basis of the interactive behavior feature combined with the specific interactive behavior information in the historical time period, the object preference distribution feature can more accurately represent the preference object of the target user in the historical time period, thereby providing more accurate and real-time object preference distribution feature for the subsequent step, that is, on the basis of improving the accuracy of object recommendation, the real-time performance of object recommendation can be further improved.

In one embodiment, as shown in FIG. 4, the interaction information includes an operation time when the interaction is triggered.

The interactive behavior information may further include an operation time when the interactive operation is triggered, based on the interactive behavior information including an object type of the interactive object that is triggered the interactive operation and object information of the interactive object. The aforementioned operation time is specifically used to indicate a timestamp of the interactive object that is triggered by the target user to perform the interactive operation. For example, interactive object B1 is triggered by the target user at time stamp C1, interactive object B2 is triggered by the target user at time stamp C2, and interactive object B3 is triggered by the target user at time stamp C3. Then the operation time included in the interactive behavior information D1 of the interactive object B1 for the target user to trigger the interactive operation is the time stamp C1, the operation time included in the interactive behavior information D2 of the interactive object B2 for the target user to trigger the interactive operation is the time stamp C2, and the operation time included in the interactive behavior information D3 of the interactive object B3 for the target user to trigger the interactive operation is the time stamp C3.

It will be appreciated that, since the one interaction described in this embodiment may be used to characterize an operation of obtaining an execution result, the one interaction may be specifically an instant interaction of obtaining the execution result, and the operation time may be a timestamp corresponding to the instant of obtaining the execution result. For example, the interactive operation is an operation of settling the commodity, and the time stamp of settling the commodity is the operation time.

The one-time interaction described in the present embodiment may also be a continuous interaction performed during the execution time interval, and then the operation time may be a start time stamp of the continuous interaction or an end time stamp of the continuous interaction. For example, the target user performs an operation of playing a single video, starts playing the video at a time stamp C1, and ends playing the video at a time stamp C2, where the operation time may be the time stamp C1 (i.e., a start time stamp) and may be the time stamp C2 (i.e., an end time stamp). Therefore, the manner of determining the operation time in the interaction behavior information is not specifically limited in this embodiment.

Based on this, determining the interaction behavior feature based on each interaction behavior information includes:

step 402, based on each operation time, ordering each interaction behavior information to obtain an interaction behavior information sequence.

The ordering mode for ordering the interaction behavior information may be: the sorting may be performed from first to second based on each operation time, or from first to second based on each operation time. Therefore, the obtained interactive behavior information sequence may be a first-to-last sequence or a second-to-first sequence. The time sequence in the specific interaction behavior information sequence needs to be specifically determined based on a sorting mode, and the specific sorting mode needs to be flexibly determined based on actual application requirements.

Specifically, the server sorts the interaction behavior information based on the operation time in the interaction behavior information to obtain an interaction behavior information sequence. For example, the operation time included in the interaction information D1 is a time stamp C1, the operation time included in the interaction information D2 is a time stamp C2, the operation time included in the interaction information D3 is a time stamp C3, and the time stamp C3 precedes the time stamp C1, and the time stamp C1 precedes the time stamp C2. If the sequence is ordered from first to last based on each operation time, the obtained interaction behavior information sequence is as follows: interactive behavior information D3, interactive behavior information D1, and interactive behavior information D2. Otherwise, if the sequence is ordered from the back to the front based on each operation time, the obtained interaction behavior information sequence is: interactive behavior information D2, interactive behavior information D1, and interactive behavior information D3.

And step 404, performing attention coding on the interaction behavior information sequence according to the time sequence to obtain the interaction behavior characteristics.

The attention code may be performed in a time sequence from first to second, or in a time sequence from second to first. Specifically, the server performs attention coding on the interaction behavior information sequence according to time sequence to obtain interaction behavior characteristics. The server performs attention coding on the interaction information sequence based on the time sequence among the interaction information in the interaction information sequence to acquire sequence modes included in the interaction information sequence and associated time sequence information among the interaction information, so that interaction behavior characteristics are obtained.

For example, if attention encoding is performed in a time sequence from first to last and the order is based on each operation time from first to last, the server will specifically use a left to right attention encoding mechanism to sequentially perform attention encoding on the interaction information D3, the interaction information D1 to the interaction information D2 in the interaction information sequence to obtain the interaction characteristics. Secondly, if attention coding is performed at a time sequence from the back to the front and sorting is performed from the front to the back based on each operation time, at this time, the server will specifically use a right-to-left attention coding mechanism to sequentially perform attention coding on the interaction information D2, the interaction information D1 and the interaction information D3 in the interaction information sequence to obtain interaction characteristics.

In this embodiment, the operation time sequences are used to ensure that operation time sequences exist among the interaction behavior information in the obtained interaction behavior information sequence, so that when the interaction behavior information in the interaction behavior information sequence is encoded according to the time sequences, the time sequence information among the interaction behavior information is further considered on the basis of obtaining the behavior information characteristics of the interaction behavior information, that is, the time sequence association included in the time sequence capturing sequence for performing the interaction operation on the interaction objects from the target user, so that the obtained interaction behavior characteristics are more accurate, more accurate interaction behavior characteristics are provided for subsequent steps, and the accuracy of object recommendation is further improved.

In one embodiment, as shown in fig. 5, the method for recommending an object further includes:

step 502, extracting object features of each interactive object according to each evaluation information.

The object features are used for characterizing object description features of the interactive object described by the target user, and the object features can at least comprise object types of the interactive object. Specifically, the server extracts object features of each interactive object through each evaluation information. Specifically, general semantic features corresponding to the evaluation information can be extracted first, and then object features can be extracted from the general semantic features corresponding to the evaluation information. As can be seen from the foregoing description, since the evaluation information includes at least: the personalized preference of the target user, the description information and the noise information of each interactive object, so that the server can determine the general semantic features corresponding to the description information of the interactive object from the general semantic features corresponding to the evaluation information, and further determine the object features of the interactive object.

Based on this, object preference distribution characteristics are generated based on object preference characteristics of the target user for each interactive object, and interactive behavior characteristics, including:

in step 504, object preference distribution features are generated based on object preference features, interaction behavior features, and object features of the target user for each interaction object.

Specifically, the server generates object preference distribution features based on object preference features, interaction behavior features, and object features of the target user for each interaction object. The server performs coding processing based on object preference characteristics, interaction behavior characteristics and object characteristics of each interaction object of the target user to obtain object preference distribution characteristics.

In order to facilitate understanding how to generate the object preference distribution feature through the object preference feature, the interaction behavior feature and the object features, as shown in fig. 6, firstly, the interaction objects triggered by the target user to perform the interaction operation in the history period are acquired, in fig. 6, n interaction objects are specifically included, and then evaluation information 601 is determined, that is, the evaluation information 601 specifically includes a plurality of evaluation information that are respectively fed back by the target user for the n interaction objects. It can then also be determined that the generated interaction behavior information 602 each time an interaction operation is triggered, i.e. the generated interaction behavior information that triggers an interaction operation for n interaction objects by the target user in the history period is specifically included in the interaction behavior information 602.

Based on this, the server extracts object features 603 of each of the n interactive objects through the evaluation information 601, extracts object preference features 604 of the target user for each of the n interactive objects through the evaluation information 601, and determines interactive behavior features 605 through the interactive behavior information 602. Thus, the server performs an encoding process based on the object feature 603, the object preference feature 604, and the interaction behavior feature 605, resulting in an object preference distribution feature.

In this embodiment, based on the object preference feature reflecting the intrinsic preference of the target user and combining the interactive behavior feature of the specific interactive behavior information in the historical time period, further from the perspective of the interactive object, the specific description information of the target user on the interactive object is represented by the object feature, and at this time, the object preference distribution feature can more accurately represent the preference object of the target user and the feature of the interactive object operated in the historical time period, so as to provide a more complete and comprehensive object preference distribution feature for the subsequent step, thereby further improving the accuracy of object recommendation.

The complete flow of how the object preference distribution feature is generated is described in detail below: in one embodiment, as shown in fig. 7, generating the object preference distribution feature based on the object preference feature, the interaction behavior feature, and the object features of the target user for each interaction object includes:

step 702, aligning the interactive behavior feature with the object preference feature of the target user for each interactive object through each object feature.

Specifically, as can be seen from the foregoing description, the interactive behavior features are at least used to characterize the behavior information features of each interactive behavior information, and the associated information features between each interactive behavior information. When the interactive behavior characteristics and the preference characteristics of each object are coded, the matching alignment of the preference characteristics of the objects of the same interactive object with the behavior information characteristics of the interactive behavior information related to the interactive object in the interactive behavior characteristics is considered. The object features are used for representing object description features of the object user describing the interactive object, so that the server can align the distribution difference between the object preference features and the interactive behavior features through the object features, namely, the server can align the object preference features of the interactive object through the object description features represented by the object features, and the behavior information features of the interactive behavior features and the interactive behavior information related to the interactive object respectively.

Illustratively, the server will specifically use a left-to-right attention encoding mechanism to sequentially perform attention encoding on the interaction information D3, the interaction information D1, and the interaction information D2 in the interaction information sequence to obtain interaction characteristics. The interaction behavior information D1 corresponds to the interaction operation triggered by the target user of the interaction object B1, the interaction behavior information D2 corresponds to the interaction operation triggered by the target user of the interaction object B2, and the interaction behavior information D3 corresponds to the interaction operation triggered by the target user of the interaction object B3.

If the evaluation information fed back by the target user for the interactive object B1, the object feature E1 and the object preference feature F1 may be extracted. The object feature E2 and the object preference feature F2 may be extracted from the evaluation information fed back by the target user for the interactive object B2. The object feature E3 and the object preference feature F3 may be extracted from the evaluation information fed back by the target user for the interactive object B3.

At this time, the server needs to align the feature extracted based on the interaction behavior information D1 from the interaction behavior features with the object preference feature F1 through the object feature E1. Similarly, the server also needs to align, through the object feature E2, the feature extracted based on the interaction information D2 from the interaction features with the object preference feature F2. And aligning the extracted feature based on the interactive behavior information D3 among the interactive behavior features with the object preference feature F3 through the object feature E3. The interaction behavior characteristic and the object preference characteristic after alignment can be specifically: the object preference feature F3, the object preference feature F1 to the object preference feature F2, and the interaction behavior feature are aligned.

And step 704, coding the aligned interactive behavior characteristics and the object preference characteristics to obtain the object preference distribution characteristics.

Specifically, the server encodes the aligned interactive behavior characteristic and the object preference characteristic to obtain the object preference distribution characteristic. For example, the interaction behavior feature and the object preference feature after alignment may be specifically: the aligned object preference feature F3, object preference feature F1 to object preference feature F2 and interaction behavior feature, i.e. the server inputs the object preference feature F3, the order of the object preference feature F1 to object preference feature F2, and the interaction behavior feature to the attention coder, and then the attention coder performs an attention coding process to output the object preference distribution feature.

It can be understood that the server may also perform encoding processing on the aligned interaction behavior feature and the object preference feature in combination with each object feature to obtain an object preference distribution feature. The attention encoder is still required to perform the attention encoding and encoding process, and the detailed description is omitted.

In this embodiment, the feature characterization of the same interactive object can be accurately encoded in the attention encoding process by aligning the object feature with the distribution difference between the object preference feature and the interactive behavior feature, so that the accuracy of the obtained object preference distribution feature is ensured, and the accuracy of object recommendation is further improved.

How the object preference features are extracted by the evaluation information, respectively, and the method of the object features will be described in detail below: in one embodiment, as shown in fig. 8, extracting, by each evaluation information, an object preference feature of a target user for each interactive object includes:

in step 802, object preference information is extracted from each of the rating information according to the object preference dimension, and each of the object preference features is determined based on each of the object preference information.

The object preference dimension is specifically a non-visual description expressed by a target user to an object, and is mainly used for expressing emotion and tendency of the object, the visual description expressed by the object is an object description dimension, and the object description dimension does not carry emotion and tendency of the object, but describes the essence of the object, so that the object description dimension can be directly object information describing the object. Based on the description information, the object preference information is the description information of emotion and tendency expressed by the target user to the object.

Specifically, the server extracts object preference information from each of the evaluation information in accordance with the object preference dimension, and determines each of the object preference characteristics based on each of the object preference information. Illustratively, taking the case of the application to the commodity recommendation scenario described in the foregoing embodiment as an example, for example, the evaluation information fed back for commodity 1 is specifically "commodity 1 size standard, but the price is somewhat expensive, and the price is hoped to be somewhat cheaper", where the server may determine that the description information for expressing emotion and tendency of the object in the evaluation information fed back for commodity 1 is specifically: "size criteria", "price is somewhat expensive" and "price is somewhat inexpensive", so the server extracts object preference information from the evaluation information in accordance with the object preference dimension specifically as follows: "size criteria", "some price is expensive", and "some price is inexpensive", and then the server can determine the object preference characteristics of the target user for the commodity 1 based on the aforementioned object preference information.

Similarly, if the evaluation information fed back for the commodity 2 is specifically "the commodity 2 is very good in light blue but is poor in quality", the server may determine that the description information for expressing emotion and tendency of the object in the evaluation information fed back for the commodity 2 is specifically: "bluish good looking" and "poor quality", so the server extracts object preference information from the evaluation information according to the object preference dimension specifically is: "bluish good looking" and "poor quality" and then the server may determine the object preference characteristics of the target user for the good 2 based on the aforementioned object preference information.

Similarly, if the evaluation information fed back for the commodity 3 is specifically "the commodity 3 is not of a proper size and is too difficult to wear", the server may determine that the description information for expressing emotion and tendency of the subject in the evaluation information fed back for the commodity 3 is specifically: the "size is not right" and "difficult to wear", so the server extracts object preference information from the evaluation information according to the object preference dimension specifically is: the "size not to" and "hard to wear" and then the server may determine the object preference characteristics of the target user for the article 3 based on the aforementioned object preference information.

Further, in the process that the server determines the preference characteristics of each object based on the preference information of each object, the server needs to specifically extract the general semantic characteristics of the evaluation information, and then determine the preference characteristics of the object of the target user for the interactive object by considering the information characteristics of the preference information of the object and the contextual information characteristics of the preference information of the object in the evaluation information through the general semantic characteristics of the evaluation information. For ease of understanding, as formula (1):

p _j ＝User-head(H _j ·W _p +b _p )； (1)

wherein p is _j Object preference feature, H, representing interaction object j _j General semantic features, W, representing rating information of interactive object j _p And b _p Are all learnable parameters.

From the above formula (1), the server may reflect the object preference taste of the target user through the object preference information in the comment information fed back by the target user from the perspective of the target user, so as to shorten the representing distance of the object preference information in the user preference layer, thereby obtaining more accurate object preference characteristics.

Based on this, object features of each interactive object are extracted by each evaluation information, including:

in step 804, object evaluation information is extracted from each evaluation information according to the object description dimension, and each object feature is determined based on each object evaluation information.

The object description dimension is specifically an intuitive description expressed by a target user to the object, namely the object description dimension does not carry emotion and tendency of the object, but describes the essence of the object, so the object description dimension can be directly object information describing the object. Based on this, the object evaluation information is the information of the object nature described by the target user, that is, the object evaluation information may be the object type and the object information, for example, the object type is sports shoes, and the object information may be the sports shoe brand and the sports shoe size.

Specifically, the server extracts object evaluation information from each evaluation information in accordance with the object description dimension, and determines each object feature based on each object evaluation information. By way of example, taking the case of the application to the commodity recommendation scenario described in the foregoing embodiment again, for example, the evaluation information fed back for commodity 1 is specifically "commodity 1 size standard, but the price is somewhat expensive, and the price is hoped to be somewhat cheaper", at this time, the description for expressing the essence of the object in the evaluation information fed back for commodity 1 may be determined by the server: "commodity 1" and commodity information related to "commodity 1", the server extracts object evaluation information from the evaluation information in accordance with the object description dimension specifically is: "commodity 1" and commodity information related to "commodity 1", and then the server may determine the object characteristics of the target user for commodity 1 based on the aforementioned object evaluation information.

Similarly, if the evaluation information fed back for the commodity 2 is "the commodity 2 is light blue and is good but of poor quality", the server may determine that the description for expressing the essence of the object in the evaluation information fed back for the commodity 2 is specifically: "commodity 2" and commodity information related to "commodity 2", the server extracts object evaluation information from the evaluation information in accordance with the object description dimension specifically is: "commodity 2" and commodity information related to "commodity 2", and then the server may determine the object characteristics of the target user for commodity 2 based on the aforementioned object evaluation information.

Similarly, if the evaluation information fed back for the commodity 3 is specifically "the commodity 3 is not of a proper size and is too difficult to be worn", the server may determine that the description for expressing the essence of the object in the evaluation information fed back for the commodity 3 is specifically: "commodity 3" and commodity information related to "commodity 3", the server extracts object evaluation information from the evaluation information in accordance with the object description dimension specifically is: "commodity 3" and commodity information related to "commodity 3", and then the server may determine the object characteristics of the target user for commodity 3 based on the aforementioned object evaluation information.

Further, in the process that the server determines the characteristics of each object based on the evaluation information of each object, the server needs to specifically extract the general semantic characteristics of the evaluation information, and then determine the characteristics of the object of the target user for the interactive object by considering the information characteristics of the evaluation information of the object and the contextual information characteristics of the evaluation information of the object in the evaluation information through the general semantic characteristics of the evaluation information. For ease of understanding, as equation (2):

v _j ＝Item-head(H _j ·W _v +b _v )； (2)

wherein v is _j Object features, H, representing interactive objects j _j General semantic features, W, representing rating information of interactive object j _V And b _V Are all learnable parameters.

From the above formula (2), the server may reflect the object features of the same interactive object from the perspective of the interactive object, so as to shorten the representing distance of the object in the object representation layer, thereby obtaining more accurate object features.

In this embodiment, from the perspective of the target user, the object preference information in the comment information fed back by the target user reflects the object preference taste of the target user, so as to obtain more accurate object preference characteristics. From the perspective of the interactive object, the comment information fed back by the same interactive object reflects the object characteristics of the same interactive object, so that more accurate object characteristics are obtained. The object preference distribution characteristics are generated by combining the more accurate object characteristics and the more accurate object preference characteristics, namely, the accuracy of determining the target object is improved, so that the accuracy of recommending the object is further improved.

In one embodiment, as shown in FIG. 9, determining a target object from among candidate objects based on object preference distribution features includes:

step 902, determining a candidate feature of each candidate in the candidate set.

Specifically, the server determines the candidate object characteristics of each candidate object in the candidate object sets, that is, the server determines the candidate object sets based on the application scene requirement first, for example, in the case of being applied to commodity recommendation scenes, the candidate object sets may be a set formed by all commodities under commodity shopping software. Alternatively, as applied to the paper text scenario, the candidate set may be a set of all paper text included in the paper text website. Based on the above, the server performs object feature extraction on each candidate object in the candidate object set, that is, the server performs feature encoding through the object type and the object information of each candidate object, so as to obtain the respective candidate object features of the candidate objects.

Illustratively, there are candidates G1, G2, G3, G4, and G5 in the candidate set, and candidate G1 corresponds to candidate feature H1, candidate G2 corresponds to candidate feature H2, candidate G3 corresponds to candidate feature H3, candidate G4 corresponds to candidate feature H4, and candidate G5 corresponds to candidate feature H5.

In step 904, respective similarities between the object preference distribution features and the candidate object features are calculated.

Wherein similarity is used to describe differences between features. Thus, the respective similarity between the object preference distribution feature and each candidate object feature is specifically used to describe: the preferred object features of the object preferred by the target user, and the differences between the candidate object features. Specifically, the server calculates the differences between the preferred object features of the objects preferred by the target user characterized by the object preference distribution features and the candidate object features to obtain the aforementioned similarity. And the similarity in this embodiment may include, but is not limited to: cosine similarity (Cosine Similarity), euclidean distance (Euclidean Distance), and the like, and the specific similarity algorithm is flexibly determined based on actual requirements.

In step 906, a target object is determined based on respective similarities between the object preference distribution feature and each candidate object feature.

Specifically, the server determines the target object based on respective similarities between the object preference distribution feature and each candidate object feature. Since the target object may be composed of one or more candidate objects, when the target object is only one candidate object, the server may consider the characteristics of the preferred object of the target user, and the similarity between the preferred object characteristics and the characteristics of each candidate object, and select, as the target object, the candidate object corresponding to the candidate object characteristic with the greatest similarity.

Illustratively, further describing the foregoing example, if the similarity between the candidate object feature H1 and the object preference distribution feature is calculated to be 40%, the similarity between the candidate object feature H2 and the object preference distribution feature is calculated to be 70%, the similarity between the candidate object feature H3 and the object preference distribution feature is calculated to be 20%, the similarity between the candidate object feature H4 and the object preference distribution feature is calculated to be 50%, and the similarity between the candidate object feature H5 and the object preference distribution feature is calculated to be 80%. At this time, 80% of the similarity, which is the largest in value, may be determined, and thus the candidate G5 corresponding to the candidate feature H5 may be determined as the target object recommended to the target user.

Optionally, determining the target object based on respective similarities between the object preference distribution feature and each candidate object feature includes: determining the similarity between the candidate object features and the object preference distribution features, wherein the candidate object features are in a similarity threshold range as target candidate object features; and determining the candidate object corresponding to the target candidate object characteristic as a target object.

The similarity threshold range may be 60% to 100%, or 70% to 100%, etc., and the similarity threshold range needs to be flexibly determined based on actual application requirements.

Specifically, after calculating the respective degrees of similarity between the object preference distribution feature and each candidate object feature, the server selects, as the target candidate object feature, a candidate object feature whose degree of similarity with the object preference distribution feature is in a similarity threshold range, where the target candidate object feature may be a single or a plurality of candidate object features, thereby determining a candidate object corresponding to the target candidate object feature as a target object, where the target object may also be composed of a single or a plurality of candidate objects.

Illustratively, further description is given with the foregoing examples, and the similarity threshold ranges from 60% to 100%. Based on this, the similarity between the candidate object feature H1 and the object preference distribution feature is calculated to be 40%, the similarity between the candidate object feature H2 and the object preference distribution feature is calculated to be 70%, the similarity between the candidate object feature H3 and the object preference distribution feature is calculated to be 20%, the similarity between the candidate object feature H4 and the object preference distribution feature is calculated to be 50%, and the similarity between the candidate object feature H5 and the object preference distribution feature is calculated to be 80%. At this time, it may be determined that the similarity 70% and the similarity 80% are in the similarity threshold range of 60% to 100%, at this time, the candidate feature H5 and the candidate feature H2 are both determined as target candidate features, and then a candidate corresponding to the target candidate feature is selected as a target object, that is, the candidate G5 corresponding to the candidate feature H5 and the candidate G2 corresponding to the candidate feature H2 are selected as target objects, that is, the target objects are specifically the candidate G2 and the candidate G5.

It will be appreciated that in practical applications, in the case where the target object is composed of a plurality of candidate objects, information such as the number of times each candidate object is triggered in a history period may also be considered, and an object that is finally recommended to the target user may be selected from the plurality of candidate objects, which will not be described in detail herein. And all examples in the present embodiment are for understanding the present scheme only, and should not be construed as specific limitations of the present scheme.

In this embodiment, by considering the respective similarity between the object preference distribution feature and each candidate object feature, the difference between each candidate object feature and the preference object feature can be reflected by the similarity, so that when the target object is determined, the candidate object with smaller difference is selected for recommendation, and the target object can be more consistent with the preference object feature, that is, the accuracy of object recommendation is more improved.

In one embodiment, as shown in FIG. 10, the process of extracting object preference features and generating object preference distribution features is implemented based on a preference prediction model.

Specifically, in the process of the foregoing embodiment, the object preference feature of the target user for each interactive object may be extracted based on the preference prediction model through each evaluation information, and the object preference distribution feature of the target user may be generated based on the object preference feature of the target user for each interactive object through the preference prediction model. In this process, the object preference distribution feature can be further generated by a preference prediction model based on the object preference feature of the target user for each interactive object and the interactive behavior feature. And sequencing the interactive behavior information based on the operation time through the preference prediction model to obtain an interactive behavior information sequence, and then performing attention coding on the interactive behavior information sequence according to the time sequence to obtain the interactive behavior characteristics.

And secondly, generating object preference distribution characteristics based on object preference characteristics, interaction behavior characteristics and object characteristics of each interaction object of the target user through a preference prediction model. And aligning the interactive behavior characteristics with the object preference characteristics of the target user aiming at each interactive object based on the object characteristics through each object characteristic through a preference prediction model, and then carrying out coding processing on the aligned interactive behavior characteristics and the object preference characteristics to obtain object preference distribution characteristics.

Accordingly, an acquisition method of the acquisition of the preference prediction model will be described in detail below: the method for acquiring the preference prediction model comprises the following steps:

step 1002, acquiring a first interaction object sample of which the target user triggers the interaction operation in a first history period, and a second interaction object sample of which the target user triggers the interaction operation in a second history period; wherein the second historical period is a next historical period adjacent to the first historical period.

Wherein the second historical period is a next historical period adjacent to the first historical period. The time intervals of the first history period and the second history period may be the same or different, and the start time stamp of the second history period may be the end time stamp of the first history period. For example, the first historical period is 2023, 7, 14, 12 to 2023, 7, 20, 12, then the second historical period may be 2023, 7, 20, 12 to 2023, 7, 26, 12.

Second, the interactive object sample may be one or more, and the interactive object sample may be merchandise, text, audio and video, virtual data, and the like. It can be seen that the interaction object sample has a corresponding relation with the operation type of the triggered interaction operation, and the operation type of the interaction operation is determined based on the application requirement. I.e. the interaction object samples are similar to the interaction objects described in the previous embodiments and will not be described here again.

Specifically, when the server needs to train the initial preference prediction model to obtain the preference prediction model, the server may respectively obtain a first interaction object sample in which the target user triggers the interaction operation in a first historical period, and a second interaction object sample in which the target user triggers the interaction operation in a second historical period. The method for acquiring the first interactive object sample and the second interactive object sample may be acquired by the server through communication connection with the terminal used by the target user, or may be acquired by the server from the data storage system, and the specific acquisition manner is similar to the method for acquiring the interactive object described in the foregoing embodiment, which is not described herein again.

In step 1004, a first evaluation information sample of each first interaction object sample of the target user is determined.

Wherein the evaluation information sample comprises at least: the personalized preference of the target user, the description information of each interactive object, and the noise information, i.e., the evaluation information sample, are similar to those described in the foregoing embodiments. The first sample of rating information is rating information fed back by the target user for the first sample of interaction objects within the first historical period of time. Specifically, the server determines a first evaluation information sample of each first interaction object sample for the target user. That is, the server acquires the first evaluation information sample in a similar manner to the acquisition of the evaluation information described in the foregoing embodiment.

In step 1006, based on each first evaluation information sample, the first object preference feature of the target user for each first interaction object sample is extracted through the initial preference prediction model.

Wherein the object preference feature is used to characterize the preferences of the target user for the interactive object. Preferences are non-intuitive descriptions, so the preferences are emotions and trends of the target user on the interactive objects, and the preferences can be specifically expressed as the preference degree of the target user on the interactive objects. Thus, the first object preference feature is used to characterize the target user's preference for the first interaction object sample over the first historical period of time. The object preference features are specifically similar to those described in the foregoing embodiments and will not be described in detail herein.

Based on the first evaluation information samples, the server inputs the first evaluation information samples into an initial preference prediction model, so that first object preference characteristics of a target user for the first interaction object samples are extracted through the initial preference prediction model. Specifically, the initial preference prediction model extracts, through a user presentation layer, first object preference characteristics of each first interaction object sample from first evaluation information samples of each first interaction object sample according to object preference dimensions. The method for extracting the object preference feature from the evaluation information sample is similar to the foregoing embodiment, and will not be repeated here.

Step 1008, generating a first object preference distribution feature of the target user based on each first object preference feature, and extracting an object attribute feature corresponding to each second interaction object sample.

The object preference distribution feature is used for characterizing a preference object feature of an object preferred by the target user, and the preference object feature may at least include an object type of the object, and the object preference distribution feature is specifically similar to that described in the foregoing embodiment and is not described herein. Thus, the first object preference feature may be used to characterize preferred object features of objects preferred by the target user over a first historical period of time. Secondly, the object attribute features are used for characterizing attribute information of the second interaction object sample, and the attribute information at least comprises: object type of the second interactive object sample, object information of the second interactive object sample, and the like.

Specifically, the server generates first object preference distribution characteristics of the target user based on the first object preference characteristics, and extracts object attribute characteristics corresponding to the second interaction object samples. Specifically, the server extracts attribute information of the second interaction object sample through the initial preference prediction model, and encodes the attribute information of the second interaction object sample to obtain object attribute characteristics. Similar to the previous embodiments, the server may, through the initial preference prediction model, attention code respective first object preference characteristics of the first interaction object samples, thereby deriving feature representation distances between the first object preference features, to generate first object preference distribution features capable of reflecting preferences of the target user over a first historical period of time.

Further, as can be seen from the foregoing embodiments, the object preference distribution feature may be obtained by taking into consideration the interaction behavior feature and the object feature. Similarly, during the training process, the server may also determine first interaction behavior information generated each time an interaction operation is triggered during a first historical period of time, and determine first interaction behavior features through an initial preference prediction model based on each first interaction behavior information. And extracting first object evaluation information from each first evaluation information sample according to the object description dimension, and determining the first object characteristics of each first interaction object sample through an initial preference prediction model based on each first object evaluation information. Thus, the server may generate the first object preference distribution feature based on each first object preference feature, each first object feature, and the first interaction behavior feature, in particular by means of an initial preference prediction model. The specific generation manner is similar to that of the foregoing embodiment, and will not be repeated here.

In step 1010, model parameters of the initial preference prediction model are adjusted by the first object preference distribution feature and each object attribute feature to obtain a preference prediction model.

Specifically, the server adjusts model parameters of the initial preference prediction model through the first object preference distribution feature and each object attribute feature to obtain a preference prediction model. The server can splice the first object preference distribution characteristics and the object attribute characteristics, input the spliced first object preference distribution characteristics and object attribute characteristics into a multi-layer perceptron in the initial preference prediction model, calculate the association degree between the first object preference distribution characteristics and the object attribute characteristics through the multi-layer perceptron, and adjust model parameters of the initial preference prediction model until the obtained association degree reaches the association degree threshold or the iteration number reaches the maximum iteration number when the association degree does not reach the association degree threshold or the iteration number does not reach the maximum iteration number, and obtain the preference prediction model based on the model parameters. And parameters of a generator in the initial preference prediction model can be adjusted through the first object preference distribution characteristics and the association degree among the object attribute characteristics, and the generator is specifically used for generating the first object preference distribution characteristics.

Based on this, the first object preference distribution feature and the degree of association between the object attribute features may be calculated as the following formula (3):

wherein D is _φ (e, x) represents the degree of association, e represents the first object preference distribution feature, x represents each object attribute feature, σ represents a sigmoid function,representing a stitching operation, MLP represents a neural network of three layers.

In order to facilitate understanding of the foregoing flow, as shown in fig. 11, first, a first interaction object sample in which the target user triggers an interaction operation in a first history period is acquired, then, a first evaluation information sample 1101 of the target user for each first interaction object sample is determined, and first interaction behavior information 1102 generated by the target user when the target user triggers an interaction operation each time in the first history period is determined.

Based on this, in a similar manner as described in the foregoing embodiment, each of the first object features 1104 and the first object preference features 1105 are extracted based on the first evaluation information sample 1101, then the first interaction behavior features 1106 are extracted based on the first interaction behavior information 1102, and the first object preference distribution features 1108 are generated by the generator 1107 based on each of the first object features 1104, the first object preference features 1105, and the first interaction behavior features 1106. At this time, each object attribute feature 1109 is also extracted based on the second interactive object sample 1103, so that the degree of association between the first object preference distribution feature 1108 and each object attribute feature 1109 can be fed back to the generator 1107 to adjust the parameters of the generator 1107, so as to guide the learning of the generator 1107 to further improve the accuracy of the object preference distribution feature.

In this embodiment, by adjusting the model parameters to obtain the preference prediction model according to the first object preference distribution feature and the association degree between the object attribute features, the obtained preference prediction model can learn how to generate the object preference distribution feature closer to the real preference from the difference between the generated object preference distribution feature and the real object attribute features, so as to improve the accuracy of the object preference distribution feature in the actual application, thereby enabling the obtained target object to better conform to the real preference of the user, and further improving the accuracy of object recommendation in the actual application.

In one embodiment, as shown in fig. 12, the method for obtaining the preference prediction model further includes:

step 1202, obtain a negative sample of the interactive object.

Wherein the negative interaction object sample is different from the second interaction object sample. Specifically, the server obtains a negative sample of the interactive object. The server may select the negative sample of the interactive object from the candidate object set by intra-batch negative sampling, or the server may select the negative sample of the interactive object from the candidate object set in consideration of the generated first object preference distribution characteristics, which is not particularly limited herein.

Based on this, by the first object preference distribution feature and each object attribute feature, model parameters of the initial preference prediction model are adjusted to obtain a preference prediction model, including:

in step 1204, model parameters of the initial preference prediction model are adjusted by performing counterlearning with the first object preference distribution feature and each object attribute feature, the interaction object negative sample, and the second interaction object sample to obtain a preference prediction model.

Specifically, the server further considers the interaction object negative sample and the second interaction object sample to perform countermeasure learning together on the basis of considering the first object preference distribution feature and each object attribute feature, so as to adjust model parameters of the initial preference prediction model, and obtain the preference prediction model. The server may splice the first object preference distribution feature and each object attribute feature, input the spliced first object preference distribution feature and each object attribute feature to a multi-layer perceptron in the initial preference prediction model, calculate the association degree between the first object preference distribution feature and each object attribute feature through the multi-layer perceptron, and then perform counterlearning with the interaction object negative sample and the second interaction object sample through the first object preference distribution feature and each object attribute feature.

Based on the above, it is known that the generator generates object preference distribution characteristics in the initial preference prediction model, and the arbiter is configured to distinguish whether the real preferences (i.e. the object attribute characteristics) reflected by the target user in the first evaluation information sample are similar to the preferences (i.e. the first object preference distribution characteristics) generated by the generator, the generator and the arbiter in the initial preference prediction model are subjected to iterative minimum training to perform countermeasure learning, and model parameters of the generator and the arbiter in the initial preference prediction model are adjusted together, so that the preference prediction model is obtained when the training completion condition is reached. It can be understood that in practical application, the application is only applied to the generator in the preference prediction model to generate the object preference distribution characteristics, and the discriminator is not used in practical application.

Further, the training completion condition may be: the loss value calculated by the first object preference distribution feature, each object attribute feature, the interaction object negative sample and the second interaction object sample reaches a loss threshold value, or the iteration number reaches the maximum iteration number. For ease of understanding, the following describes how to iterate the method of performing the minimum maximum training for countermeasure learning, taking equation (4) as an example:

/>

Wherein e ^t ＝G(H _(n+1) ；S _u ) Representing attribute characteristics of each object, e ^g ＝G(P；S _u ) Representing a first object preference distribution feature, x' _j Representing a negative sample of the interactive object, x _j ～p _pos ,x′ _j ～p _neg Representing a second interaction object sample.

In order to facilitate understanding of the foregoing flow, as shown in fig. 13, first, a first interaction object sample in which the target user triggers an interaction operation in a first history period is acquired, then, a first evaluation information sample 1301 of the target user for each first interaction object sample is determined, and first interaction behavior information 1302 generated by the target user each time the target user triggers an interaction operation in the first history period is determined.

Based on this, in a similar manner as described in the foregoing embodiment, each of the first object features 1304 and the first object preference features 1305 are extracted based on the first evaluation information sample 1301, then the first interaction behavior features 1306 are extracted based on the first interaction behavior information 1302, and the first object preference distribution features 1308 are generated by the generator 1307 based on each of the first object features 1304, the first object preference features 1305, and the first interaction behavior features 1306. At this time, the parameters of the generator 1307 and the discriminator 1311 are also adjusted by extracting the object attribute features 1309 based on the second interactive object sample 1303, and then jointly adjusting the parameters of the generator 1307 and the discriminator 1311 by the first object preference distribution feature 1308 and the object attribute features 1309, and the second interactive object sample 1303 and the interactive object negative sample 1310. It is to be understood that the foregoing examples have been provided merely for the understanding of the present invention and are not to be construed as limiting the invention in any way.

Further, in practical application, a cross entropy loss function may be introduced to optimize the object recommendation capability, that is, calculate a cross entropy loss value between the negative sample of the interactive object and the second sample of the interactive object, so as to adjust the process of determining the target object through the cross entropy loss value, specifically as shown in formula (5):

wherein,representing cross entropy loss value, x' _j A negative sample of the interactive object is represented,x _j ～p _pos ,x′ _j ～p _neg representing a second interaction object sample.

In this embodiment, through the first object preference distribution feature and the association degree between the object attribute features, further through introducing the interaction object negative sample and the second interaction object sample to perform countermeasure learning, the preference prediction model can be guided to learn how to generate the object preference distribution feature closer to the real preference, so as to further improve the accuracy of object recommendation in practical application.

The method of how to obtain negative examples of interactive objects will be described in detail as follows: in one embodiment, as shown in FIG. 14, obtaining a negative sample of an interactive object includes:

step 1402, performing intra-batch negative sampling on a plurality of candidate objects in the candidate object set, and determining the candidate objects obtained by the intra-batch negative sampling as negative samples of the interaction objects.

Specifically, the server performs intra-batch (batch) negative sampling on a plurality of candidate objects in the candidate object set, and determines the candidate objects obtained by the intra-batch negative sampling as interaction object negative samples. That is, the server selects a candidate object different from the second interaction object sample from the plurality of candidate objects at each iteration, and then determines the candidate object different from the second interaction object sample as the interaction object negative sample for the iteration.

Or, in another embodiment, obtaining a negative sample of the interactive object includes:

step 1404, calculating respective similarities between the first object preference distribution feature and respective candidate object features of the respective candidate objects.

Specifically, the server calculates respective similarities between the first object preference distribution feature and respective candidate object features of the respective candidate objects. The method for calculating the similarity is similar to the method for calculating the respective similarity between the object preference distribution feature and each candidate object feature described in the foregoing embodiment, and will not be described here.

In step 1406, candidate object features that are not within a similarity threshold range are determined to be interaction candidate object features, with the similarity between the first object preference distribution features.

Specifically, the server determines candidate object features, which are not within a similarity threshold range, as interaction candidate object features, of similarity with the first object preference distribution feature. For example, describing 60% to 100%, if the similarity between the candidate object feature H1 and the first object preference distribution feature is calculated to be 40%, the similarity between the candidate object feature H2 and the first object preference distribution feature is calculated to be 70%, and the similarity between the candidate object feature H3 and the first object preference distribution feature is calculated to be 20%, the similarity that is not in the range of 60% to 100% is calculated to be: 20% and 40%, candidate feature H1 and candidate feature H3 can thus be determined as interaction candidate features.

At step 1408, the candidate object corresponding to the interaction candidate object feature is determined to be an interaction object negative sample.

Specifically, a candidate object corresponding to the interaction candidate object feature is determined as an interaction object negative sample. For example, further to the foregoing example, the candidate feature H1 and the candidate feature H3 are determined as the interaction candidate feature, that is, the candidate G1 corresponding to the candidate feature H1 and the candidate G3 corresponding to the candidate feature H3 are taken as the interaction object negative samples.

Further, since the negative interaction object sample is different from the second interaction object sample, the server may determine the candidate object corresponding to the feature of the interaction candidate object as the negative interaction object sample, and then select the negative interaction object sample different from the second interaction object sample as the negative interaction object sample. Not described in detail here.

In this embodiment, the negative sample of the interactive object is obtained in different modes, so that the flexibility of the model training process can be improved, and the diversity of the obtained sample can be improved, so that the effect of antagonism learning is improved, the reliability of the preference prediction model is improved, and the accuracy of object recommendation in practical application is further improved.

Based on the foregoing detailed description of the embodiments, a complete flow of the method for object recommendation in the embodiments of the present application will be described, and in one embodiment, as shown in fig. 15, a method for object recommendation is provided, which is illustrated by using the server 104 in fig. 1 as an example, where it is understood that the method may also be applied to the terminal 102, and may also be applied to a system including the terminal 102 and the server 104, and implemented through interaction between the terminal 102 and the server 104. In this embodiment, the method includes the steps of:

Step 1501, obtain the interactive objects triggered by the target user to perform interactive operation in the historical time period, and determine the evaluation information fed back by the target user for each interactive object respectively.

The interactive objects can be one or more, and the interactive objects can be commodities, texts, audios and videos, virtual data and the like. It can be seen that the interactive object has a corresponding relationship with the operation type of the triggered interactive operation, and the operation type of the interactive operation is determined based on the application requirement.

Next, the evaluation information includes at least: personalized preferences of the target user, descriptive information of each interactive object, and noise information. And the evaluation information may include data information in text format, may also include data information in picture format, or even data information in audio format. It is to be understood that in practical application, the situation that the evaluation information of the target user for a certain interaction object is empty may also occur, which is not limited herein.

Specifically, in a scenario in which object recommendation is performed for a target user, a server can acquire an interactive object in which an interactive operation is triggered by the target user in a history period. That is, the server may obtain, from the data information stored in the terminal, an interactive object that is triggered by the target user to perform an interactive operation during the history period through a communication connection with the terminal used by the target user. Or after each user triggers the interactive operation on each object in the historical time period, the interactive object of each user can be transmitted to the server in real time, and the server stores the interactive object of each user in the data storage system, so that when object recommendation is required for the target user, the interactive object of each user, which is triggered by the target user in the historical time period, is determined from the interactive objects of each user stored in the data storage system.

Further, the server further acquires evaluation information fed back by the target user for each interactive object. That is, when the target user feeds back for the interactive object, the server may acquire evaluation information including various information. Secondly, the data information included in the evaluation information may be null, and the target user may not perform feedback evaluation on the interaction object, and at this time, the evaluation information acquired by the server is null.

In step 1502, object preference information is extracted from each of the rating information according to the object preference dimensions, and each of the object preference features is determined based on each of the object preference information.

The object preference dimension is specifically a non-visual description expressed by a target user to an object, and is mainly used for expressing emotion and tendency of the object, the visual description expressed by the object is an object description dimension, and the object description dimension does not carry emotion and tendency of the object, but describes the essence of the object, so that the object description dimension can be directly object information describing the object. Based on the description information, the object preference information is the description information of emotion and tendency expressed by the target user to the object. Specifically, the server extracts object preference information from each of the evaluation information in accordance with the object preference dimension, and determines each of the object preference characteristics based on each of the object preference information.

In step 1503, object evaluation information is extracted from each evaluation information according to the object description dimension, and each object feature is determined based on each object evaluation information.

The object description dimension is specifically an intuitive description expressed by a target user to the object, namely the object description dimension does not carry emotion and tendency of the object, but describes the essence of the object, so the object description dimension can be directly object information describing the object. Based on this, the object evaluation information is the information of the object nature described by the target user, that is, the object evaluation information may be the object type and the object information, for example, the object type is sports shoes, and the object information may be the sports shoe brand and the sports shoe size. Specifically, the server extracts object evaluation information from each evaluation information in accordance with the object description dimension, and determines each object feature based on each object evaluation information.

In step 1504, interaction behavior information generated each time an interaction is triggered is determined.

Wherein, the interaction behavior information at least comprises: the object type of the interactive object to be triggered by the interactive operation, the object information of the interactive object, the operation time when the interactive operation is triggered, and the like.

Step 1505, based on each operation time, ordering each interaction behavior information to obtain an interaction behavior information sequence.

The ordering mode for ordering the interaction behavior information may be: the sorting may be performed from first to second based on each operation time, or from first to second based on each operation time. Therefore, the obtained interactive behavior information sequence may be a first-to-last sequence or a second-to-first sequence. The time sequence in the specific interaction behavior information sequence needs to be specifically determined based on a sorting mode, and the specific sorting mode needs to be flexibly determined based on actual application requirements. Specifically, the server sorts the interaction behavior information based on the operation time in the interaction behavior information to obtain an interaction behavior information sequence.

And step 1506, performing attention coding on the interactive behavior information sequence according to the time sequence to obtain the interactive behavior characteristics.

The attention code may be performed in a time sequence from first to second, or in a time sequence from second to first. And secondly, the interactive behavior features are at least used for representing the behavior information features of each interactive behavior information and the associated information features among the interactive behavior information.

Specifically, the server performs attention coding on the interaction behavior information sequence according to time sequence to obtain interaction behavior characteristics. The server performs attention coding on the interaction information sequence based on the time sequence among the interaction information in the interaction information sequence to acquire sequence modes included in the interaction information sequence and associated time sequence information among the interaction information, so that interaction behavior characteristics are obtained.

In step 1507, the interaction behavior features are aligned with the object preference features of the target user for each interaction object by each object feature.

And step 1508, performing coding processing on the aligned interactive behavior characteristics and the object preference characteristics to obtain object preference distribution characteristics.

Specifically, the server encodes the aligned interactive behavior characteristic and the object preference characteristic to obtain the object preference distribution characteristic. It can be understood that the server may also perform encoding processing on the aligned interaction behavior feature and the object preference feature in combination with each object feature to obtain an object preference distribution feature. The attention encoder is still required to perform the attention encoding and encoding process, and the detailed description is omitted.

In step 1509, the candidate features of each candidate in the candidate set are determined, and the respective similarity between the object preference distribution feature and each candidate feature is calculated.

Wherein similarity is used to describe differences between features. Thus, the respective similarity between the object preference distribution feature and each candidate object feature is specifically used to describe: the preferred object features of the object preferred by the target user, and the differences between the candidate object features.

Specifically, the server determines the candidate object characteristics of each candidate object in the candidate object sets, that is, the server determines the candidate object sets based on the application scene requirement first, for example, in the case of being applied to commodity recommendation scenes, the candidate object sets may be a set formed by all commodities under commodity shopping software. Alternatively, as applied to the paper text scenario, the candidate set may be a set of all paper text included in the paper text website. Based on the above, the server performs object feature extraction on each candidate object in the candidate object set, that is, the server performs feature encoding through the object type and the object information of each candidate object, so as to obtain the respective candidate object features of the candidate objects. The server then calculates the differences between the preferred object features of the objects preferred by the target user, characterized by the object preference distribution features, and the candidate object features to obtain the aforementioned similarities.

In step 1510, the candidate object features within the similarity threshold range are determined as target candidate object features, and the candidate object corresponding to the target candidate object features is determined as target object.

Step 1511, recommending the target object to the target user.

Specifically, the server recommends the determined target object to the target user. In this embodiment, the server 104 in fig. 1 is taken as an example, so that the server may send the determined target object to the terminal through a communication connection with the terminal used by the target user, so that the terminal used by the target user recommends the target object.

It should be understood that the specific implementation of steps 1501 to 1515 is similar to the previous embodiment, and will not be repeated here.

It should be understood that, although the steps in the flowcharts related to the embodiments described above are sequentially shown as indicated by arrows, these steps are not necessarily sequentially performed in the order indicated by the arrows. The steps are not strictly limited to the order of execution unless explicitly recited herein, and the steps may be executed in other orders. Moreover, at least some of the steps in the flowcharts described in the above embodiments may include a plurality of steps or a plurality of stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of the steps or stages is not necessarily performed sequentially, but may be performed alternately or alternately with at least some of the other steps or stages.

Based on the same inventive concept, the embodiment of the application also provides an object recommendation device for realizing the above-mentioned object recommendation method. The implementation of the solution provided by the device is similar to the implementation described in the above method, so the specific limitation in the embodiments of the one or more object recommendation devices provided below may refer to the limitation of the object recommendation method described above, and will not be repeated here.

In one embodiment, as shown in fig. 16, there is provided an object recommendation apparatus including: an evaluation information determination module 1602, an object preference feature extraction module 1604, an object preference distribution feature generation module 1606, and an object recommendation module 1608, wherein:

the evaluation information determining module 1602 is configured to obtain interactive objects triggered by the target user to perform interactive operation in a historical time period, and determine evaluation information fed back by the target user for each interactive object respectively;

the object preference feature extraction module 1604 is configured to extract, according to each piece of evaluation information, object preference features of the target user for each interaction object, where the object preference features are used to characterize preferences of the target user for the interaction object;

an object preference distribution feature generation module 1606 for generating object preference distribution features of the target user based on the object preference features of the target user for each interactive object, the object preference distribution features being used to characterize preferred object features of the objects preferred by the target user;

An object recommendation module 1608 for determining a target object from the candidate objects based on the object preference distribution characteristics and recommending the target object to the target user.

In one embodiment, as shown in fig. 17, the object recommendation apparatus further includes an interactive behavior feature determination module 1702;

the interactive behavior feature determining module 1702 is configured to determine interactive behavior information generated when the interactive operation is triggered each time, and determine interactive behavior features based on each interactive behavior information;

the object preference distribution feature generating module 1606 is specifically configured to generate object preference distribution features based on the object preference features of the target user for each interactive object and the interactive behavior features.

In one embodiment, the interaction behavior information includes an operation time when the interaction operation is triggered;

the interactive behavior feature determining module 1702 is specifically configured to sort, based on each operation time, each interactive behavior information to obtain an interactive behavior information sequence; and carrying out attention coding on the interaction behavior information sequence according to the time sequence to obtain the interaction behavior characteristics.

In one embodiment, as shown in fig. 18, the object recommendation apparatus further includes an object feature extraction module 1802;

the object feature extraction module 1802 is configured to extract object features of each interactive object through each evaluation information;

The object preference distribution feature generating module 1606 is specifically configured to generate an object preference distribution feature based on the object preference feature, the interaction behavior feature, and the object feature of each interaction object of the target user.

In one embodiment, the object preference distribution feature generation module 1606 is specifically configured to align, through each object feature, the interaction behavior feature with the object preference feature of the target user for each interaction object; and carrying out coding processing on the aligned interactive behavior characteristics and the object preference characteristics to obtain the object preference distribution characteristics.

In one embodiment, the object preference feature extraction module 1604 is specifically configured to extract object preference information from each of the evaluation information according to the object preference dimension, and determine each of the object preference features based on each of the object preference information;

the object feature extraction module 1802 is specifically configured to extract object evaluation information from each piece of evaluation information according to an object description dimension, and determine each object feature based on each piece of object evaluation information.

In one embodiment, the object recommendation module 1608 is specifically configured to determine candidate object features of each candidate object in the candidate object set; calculating respective similarity between the object preference distribution characteristics and each candidate object characteristic; the target object is determined based on respective similarities between the object preference distribution feature and each candidate object feature.

In one embodiment, the object recommendation module 1608 is specifically configured to determine, as the target candidate object feature, a candidate object feature having a similarity with the object preference distribution feature and within a similarity threshold; and determining the candidate object corresponding to the target candidate object characteristic as a target object.

In one embodiment, the process of extracting object preference features and generating object preference distribution features is implemented based on a preference prediction model;

as shown in fig. 19, the object recommendation apparatus further includes an acquisition module 1902 of a preference prediction model;

an acquisition module 1902 of a preference prediction model, configured to acquire a first interaction object sample of an interaction operation triggered by a target user in a first history period, and a second interaction object sample of an interaction operation triggered by the target user in a second history period; wherein the second historical time period is a next historical time period adjacent to the first historical time period; determining a first evaluation information sample of each first interaction object sample aiming at a target user; extracting first object preference characteristics of a target user aiming at each first interaction object sample based on each first evaluation information sample through an initial preference prediction model, and extracting object attribute characteristics corresponding to each second interaction object sample; generating first object preference distribution characteristics of the target user based on the first object preference characteristics, and extracting object attribute characteristics corresponding to each second interaction object sample; and adjusting model parameters of the initial preference prediction model through the first object preference distribution characteristics and the object attribute characteristics to obtain the preference prediction model.

In one embodiment, the obtaining module 1902 of the preference prediction model is further configured to obtain a negative sample of the interaction object; model parameters of the initial preference prediction model are adjusted through the first object preference distribution feature, each object attribute feature, the negative sample of the interactive object and the second interactive object sample for antagonism learning so as to obtain the preference prediction model

In one embodiment, the obtaining module 1902 of the preference prediction model is further configured to perform intra-batch negative sampling on a plurality of candidate objects in the candidate object set, and determine a candidate object obtained by the intra-batch negative sampling as an interaction object negative sample.

In one embodiment, the obtaining module 1902 of the preference prediction model is further configured to calculate a first object preference distribution feature and a respective similarity between the first object preference distribution feature and a respective candidate object feature of each candidate object; determining candidate object features which are not in a similarity threshold range as interaction candidate object features according to the similarity between the candidate object features and the first object preference distribution features; and determining the candidate object corresponding to the interaction candidate object characteristic as an interaction object negative sample.

The modules in the object recommendation apparatus may be implemented in whole or in part by software, hardware, or a combination thereof. The above modules may be embedded in hardware or may be independent of a processor in the computer device, or may be stored in software in a memory in the computer device, so that the processor may call and execute operations corresponding to the above modules.

In one embodiment, a computer device is provided, which may be a server or a terminal, and in this embodiment, the computer device is taken as a server to be described as an example, and the internal structure thereof may be as shown in fig. 20. The computer device includes a processor, a memory, an Input/Output interface (I/O) and a communication interface. The processor, the memory and the input/output interface are connected through a system bus, and the communication interface is connected to the system bus through the input/output interface. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database of the computer device is used for storing data related to the embodiment of the application, such as interactive objects, object prediction models and the like. The input/output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to implement a method of object recommendation.

It will be appreciated by those skilled in the art that the structure shown in fig. 20 is merely a block diagram of a portion of the structure associated with the present application and is not limiting of the computer device to which the present application applies, and that a particular computer device may include more or fewer components than shown, or may combine some of the components, or have a different arrangement of components.

In an embodiment, there is also provided a computer device comprising a memory and a processor, the memory having stored therein a computer program, the processor implementing the steps of the method embodiments described above when the computer program is executed.

In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored which, when executed by a processor, carries out the steps of the method embodiments described above.

In an embodiment, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the steps of the method embodiments described above.

It should be noted that, the object information (including, but not limited to, object device information, object personal information, etc.) and the data (including, but not limited to, data for analysis, stored data, presented data, etc.) related to the present application are both information and data authorized by the object or sufficiently authorized by each party, and the collection, use, and processing of the related data are required to comply with the related laws and regulations and standards of the related countries and regions.

Those skilled in the art will appreciate that implementing all or part of the above described methods may be accomplished by way of a computer program stored on a non-transitory computer readable storage medium, which when executed, may comprise the steps of the embodiments of the methods described above. Any reference to memory, database, or other medium used in the various embodiments provided herein may include at least one of non-volatile and volatile memory. The nonvolatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, high density embedded nonvolatile Memory, resistive random access Memory (ReRAM), magnetic random access Memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric Memory (Ferroelectric Random Access Memory, FRAM), phase change Memory (Phase Change Memory, PCM), graphene Memory, and the like. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, and the like. By way of illustration, and not limitation, RAM can be in the form of a variety of forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), and the like. The databases referred to in the various embodiments provided herein may include at least one of relational databases and non-relational databases. The non-relational database may include, but is not limited to, a blockchain-based distributed database, and the like. The processors referred to in the embodiments provided herein may be general purpose processors, central processing units, graphics processors, digital signal processors, programmable logic units, quantum computing-based data processing logic units, etc., without being limited thereto.

The technical feature information of the above embodiments may be arbitrarily combined, and for brevity of description, all possible combinations of the technical feature information in the above embodiments are not described, however, as long as there is no contradiction between the combinations of the technical feature information, they should be considered as the scope of the description.

The above examples only represent a few embodiments of the present application, which are described in more detail and are not to be construed as limiting the scope of the present application. It should be noted that it would be apparent to those skilled in the art that various modifications and improvements could be made without departing from the spirit of the present application, which would be within the scope of the present application. Accordingly, the scope of protection of the present application shall be subject to the appended claims.

Claims

1. A method of object recommendation, comprising:

Generating object preference distribution characteristics of the target user for each interactive object based on the object preference characteristics of the target user, wherein the object preference distribution characteristics are used for representing preference object characteristics of objects preferred by the target user;

and determining a target object from the candidate objects based on the object preference distribution characteristics, and recommending the target object to the target user.

2. The method according to claim 1, wherein the method further comprises:

determining interaction behavior information generated when the interaction operation is triggered each time, and determining interaction behavior characteristics based on each interaction behavior information;

the generating the object preference distribution feature of the target user based on the object preference feature of the target user for each interactive object includes:

and generating object preference distribution characteristics based on the object preference characteristics of the target user for each interactive object and the interactive behavior characteristics.

3. The method according to claim 2, wherein the interactive behavior information includes an operation time when the interactive operation is triggered;

the determining the interactive behavior feature based on each piece of interactive behavior information comprises the following steps:

Based on the operation time, ordering the interactive behavior information to obtain an interactive behavior information sequence;

and carrying out attention coding on the interactive behavior information sequence according to the time sequence to obtain the interactive behavior characteristics.

4. The method according to claim 2, wherein the method further comprises:

extracting object characteristics of each interaction object through each evaluation information;

the generating object preference distribution features based on the object preference features of the target user for each of the interactive objects and the interactive behavior features includes:

an object preference distribution feature is generated based on the object preference feature, the interaction behavior feature, and the object features of the target user for each of the interaction objects.

5. The method of claim 4, wherein the generating an object preference distribution feature based on the object preference feature, the interaction behavior feature, and the object features of the target user for each of the interaction objects comprises:

aligning the interactive behavior features with object preference features of the target user for each interactive object through each object feature;

And carrying out coding processing on the aligned interactive behavior characteristic and the object preference characteristic to obtain an object preference distribution characteristic.

6. The method according to claim 4, wherein extracting object preference characteristics of the target user for each of the interactive objects through each of the evaluation information includes:

extracting object preference information from each piece of evaluation information according to object preference dimensions, and determining each piece of object preference characteristics based on each piece of object preference information;

the extracting the object feature of each interactive object through each evaluation information includes:

object evaluation information is extracted from each evaluation information according to an object description dimension, and each object feature is determined based on each object evaluation information.

7. The method of claim 1, wherein said determining a target object from each of said candidate objects based on said object preference distribution characteristics comprises:

determining the candidate object characteristics of each candidate object in the candidate object set;

calculating respective similarity between the object preference distribution feature and each candidate object feature;

a target object is determined based on respective similarities between the object preference distribution feature and each of the candidate object features.

8. The method of claim 7, wherein the determining the target object based on respective similarities between the object preference distribution feature and each of the candidate object features comprises:

determining the similarity between the candidate object features and the object preference distribution features, wherein the candidate object features in the similarity threshold range are target candidate object features;

and determining the candidate object corresponding to the target candidate object characteristic as the target object.

9. The method of claim 1, wherein the process of extracting the object preference feature and generating the object preference distribution feature is implemented based on a preference prediction model;

the method for acquiring the preference prediction model comprises the following steps:

acquiring a first interaction object sample of which the target user triggers the interaction operation in a first historical time period and a second interaction object sample of which the target user triggers the interaction operation in a second historical time period; wherein the second historical time period is a next historical time period adjacent to the first historical time period;

determining respective first evaluation information samples of the target user aiming at each first interaction object sample;

Extracting respective first object preference characteristics of the target user for each first interaction object sample based on each first evaluation information sample through an initial preference prediction model;

generating first object preference distribution characteristics of the target user based on the first object preference characteristics, and extracting object attribute characteristics corresponding to the second interaction object samples;

and adjusting model parameters of the initial preference prediction model through the first object preference distribution characteristics and each object attribute characteristic to obtain the preference prediction model.

10. The method according to claim 9, wherein the method further comprises:

acquiring a negative sample of the interactive object;

the adjusting the model parameters of the initial preference prediction model through the first object preference distribution feature and each object attribute feature to obtain the preference prediction model includes:

and adjusting model parameters of the initial preference prediction model through the first object preference distribution characteristics, the object attribute characteristics, the interaction object negative samples and the second interaction object samples to obtain the preference prediction model.

11. The method of claim 10, wherein the obtaining a negative sample of the interactive object comprises:

performing intra-batch negative sampling on a plurality of candidate objects in a candidate object set, and determining the candidate objects obtained by the intra-batch negative sampling as the interaction object negative samples;

or alternatively, the first and second heat exchangers may be,

calculating respective similarity between the first object preference distribution feature and respective candidate object features of each candidate object;

determining the candidate object features which are not in the similarity threshold range as interaction candidate object features according to the similarity between the candidate object features and the first object preference distribution features;

and determining the candidate object corresponding to the interaction candidate object characteristic as the interaction object negative sample.

12. An object recommendation device, the device comprising:

the evaluation information determining module is used for acquiring interactive objects triggered by a target user to perform interactive operation in a historical time period and determining evaluation information fed back by the target user for each interactive object respectively;

the object preference feature extraction module is used for extracting object preference features of the target user for the interactive objects through the evaluation information, and the object preference features are used for representing the preference of the target user for the interactive objects;

An object preference distribution feature generation module, configured to generate, based on object preference features of the target user for each of the interactive objects, object preference distribution features of the target user, where the object preference distribution features are used to characterize preferred object features of objects preferred by the target user;

and the object recommendation module is used for determining a target object from the candidate objects based on the object preference distribution characteristics and recommending the target object to the target user.

13. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any one of claims 1 to 11 when the computer program is executed.

14. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any of claims 1 to 11.

15. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any one of claims 1 to 11.