CN111966914B

CN111966914B - Content recommendation method and device based on artificial intelligence and computer equipment

Info

Publication number: CN111966914B
Application number: CN202011152259.XA
Authority: CN
Inventors: 朱勇椿; 葛凯凯; 张旭; 林乐宇; 庄福振
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2020-10-26
Filing date: 2020-10-26
Publication date: 2021-01-15
Anticipated expiration: 2040-10-26
Also published as: CN111966914A

Abstract

The application relates to a content recommendation method and device based on artificial intelligence and computer equipment. The method comprises the following steps: acquiring source domain user characteristics corresponding to a new user of a target domain in a source domain; performing feature mapping on the source domain user features through a trained task-oriented mapping model to obtain mapping user features mapped from the source domain to the target domain; through a pre-trained content recommendation model, click prediction of recommended content is carried out according to the mapping user characteristics and the target domain content characteristics of each candidate content in the target domain; screening contents to be recommended from the candidate contents according to a click prediction result; and pushing the content to be recommended to the new user of the target domain. By adopting the method, the accuracy of content recommendation for the new user can be effectively improved.

Description

Content recommendation method and device based on artificial intelligence and computer equipment

Technical Field

The present application relates to the field of internet technologies, and in particular, to a content recommendation method and apparatus based on artificial intelligence, a computer device, and a storage medium.

Background

With the rapid development of internet technology, recommendation technologies based on big data and artificial intelligence have come into existence, and many recommendation systems are available, which can intelligently recommend contents in which users are interested to users. The recommendation system can generally predict the preference of the user by analyzing the historical behavior information of the user, so as to recommend information which is interested by the user, and meet the personalized recommendation requirement of the user.

In a recommendation system, for a new user who does not have history data in the field, it is a common practice to recommend information with a relatively high degree of popularity to the new user according to the degree of popularity ranking of the information. However, in the conventional method, due to lack of user behavior characteristics of the new user, it is difficult to predict the interest of the user, and personalized content recommendation cannot be accurately performed on the new user, so that the accuracy of information recommendation is not high.

Disclosure of Invention

In view of the foregoing, it is desirable to provide a content recommendation method, device, computer device and storage medium based on artificial intelligence, which can effectively improve the accuracy of content recommendation for new users.

A method of artificial intelligence based content recommendation, the method comprising:

acquiring source domain user characteristics corresponding to a new user of a target domain in a source domain;

performing feature mapping on the source domain user features through a trained task-oriented mapping model to obtain mapping user features mapped from the source domain to the target domain;

through a pre-trained content recommendation model, click prediction of recommended content is carried out according to the mapping user characteristics and the target domain content characteristics of each candidate content in the target domain;

screening contents to be recommended from the candidate contents according to a click prediction result;

and pushing the content to be recommended to the new user of the target domain.

A method of artificial intelligence based content processing, the method comprising:

acquiring historical interaction information and a corresponding click record label; the historical interaction information comprises sample source domain user characteristics of a historical user in a source domain and sample target domain content characteristics in a target domain;

performing feature mapping on the sample source domain user features through a task-oriented mapping model to be trained to obtain sample mapping user features mapped from the source domain to the target domain;

inputting the sample mapping user characteristics and the sample target domain content characteristics into a pre-trained content recommendation model, and performing click prediction on sample target domain content;

determining a task guide loss value according to a click prediction result and the click record label;

and adjusting parameters of the task guide mapping model according to the task guide loss value and continuing training until the training stopping condition is met.

An artificial intelligence based content recommendation apparatus, the apparatus comprising:

the data acquisition module is used for acquiring source domain user characteristics corresponding to a new user of a target domain in a source domain;

the feature mapping module is used for performing feature mapping on the source domain user features through a trained task-oriented mapping model to obtain mapping user features which are mapped from the source domain to the target domain;

the content prediction module is used for performing click prediction on recommended content according to the mapping user characteristics and the target domain content characteristics of each candidate content in the target domain through a pre-trained content recommendation model;

the content recommending module is used for screening the content to be recommended from the candidate content according to the click prediction result; and pushing the content to be recommended to the new user of the target domain.

In one embodiment, the data obtaining module is further configured to search, in a plurality of source domains, a source domain having interaction behavior information corresponding to the user identifier according to the user identifier of the new user in the target domain; acquiring source domain user characteristics corresponding to the user identification from the source domain; the source domain user characteristics are obtained by performing characteristic extraction based on the interactive behavior information of the user identification in the source domain.

In one embodiment, the data obtaining module is further configured to obtain an access request corresponding to a new user in the first sub-application, where the access request includes a user identifier; the first sub-application corresponds to the target domain; the first sub-application runs in an environment provided by a parent application; searching a second sub-application running in the parent application and having interactive behavior information corresponding to the user identifier according to the user identifier; the second sub-application corresponds to the source domain; and the second sub-application comprises source domain user characteristics obtained based on the interactive behavior information of the user identification.

In one embodiment, the content recommendation model is obtained by a first-stage training step, and the device further includes a first model training module for obtaining historical interaction information and corresponding click record labels in the target domain; the historical interaction information comprises sample target domain user characteristics and sample target domain content characteristics corresponding to historical users in the target domain; the click record label is a click record of sample target domain content in the target domain; inputting the sample target domain user characteristics and the sample target domain content characteristics into a content recommendation model to be trained, and performing click prediction on sample target domain content; determining a click prediction loss value according to a click prediction result and the click record label; and adjusting parameters of the content recommendation model according to the click prediction loss value and continuing training until the training stopping condition is met.

In one embodiment, the task-oriented mapping model is obtained through training in a second-stage training step, and the device further comprises a second model training module for obtaining historical interaction information and corresponding click record labels; the historical interaction information comprises sample source domain user characteristics of a historical user in a source domain and sample target domain content characteristics in a target domain; performing feature mapping on the sample source domain user features through a task-oriented mapping model to be trained to obtain sample mapping user features mapped from the source domain to the target domain; inputting the sample mapping user characteristics and the sample target domain content characteristics into a pre-trained content recommendation model, and performing click prediction on sample target domain content; determining a task guide loss value according to a click prediction result and the click record label; and adjusting parameters of the task guide mapping model according to the task guide loss value and continuing training until the training stopping condition is met.

An artificial intelligence based content processing apparatus, the apparatus comprising:

the sample acquisition module is used for acquiring historical interaction information and corresponding click record labels; the historical interaction information comprises sample source domain user characteristics of a historical user in a source domain and sample target domain content characteristics in a target domain;

the characteristic mapping module is used for carrying out characteristic mapping on the sample source domain user characteristics through a task-oriented mapping model to be trained to obtain sample mapping user characteristics which are mapped from the source domain to the target domain;

the content prediction module is used for inputting the sample mapping user characteristics and the sample target domain content characteristics into a pre-trained content recommendation model to perform click prediction on sample target domain content;

the parameter adjusting module is used for determining a task guide loss value according to a click prediction result and the click record label; and adjusting parameters of the task guide mapping model according to the task guide loss value and continuing training until the training stopping condition is met.

In one embodiment, the sample obtaining module is further configured to obtain user identifiers corresponding to historical users shared in the source domain and the target domain; respectively acquiring historical interaction information of the historical user in the source domain and the target domain according to the user identification; extracting sample source domain user characteristics of the historical users in the source domain, sample target domain content characteristics of the historical users in the target domain and click record labels corresponding to the historical interaction information in the target domain from the historical interaction information; and the click record label is a click record corresponding to each sample target domain content in the historical interaction information of the target domain.

In one embodiment, the feature mapping module is further configured to input the sample source-domain user features into a task-oriented mapping model to be trained, and perform feature matrix transformation on the sample source-domain user features through a mapping layer of the task-oriented mapping model; and outputting sample mapping user characteristics for mapping the sample source domain user characteristics to the target domain according to the result of the characteristic matrix transformation.

In one embodiment, the content prediction module is further configured to input the sample mapping user characteristics and the sample target domain content characteristics into a pre-trained content recommendation model, and predict, through the content recommendation model, a sample interest degree of the historical user in each sample target domain content; and performing click prediction on the content of each sample target domain according to the sample interest degree to obtain a click prediction result corresponding to the content of each sample target domain.

In one embodiment, the parameter adjusting module is further configured to determine a task oriented loss value according to a difference between a result of the click prediction and the click record tag; and adjusting parameters of the task guide mapping model according to the task guide loss value so as to reduce the difference between the click prediction result and the click record label in the iterative training process of the task guide mapping model.

In one embodiment, the content recommendation model is obtained through training in a first-stage training step, the task guide mapping model is obtained through training in a second-stage training step, and the device further comprises a first model training module for obtaining historical interaction information and corresponding click record labels in the target domain; the historical interaction information comprises sample target domain user characteristics and sample target domain content characteristics corresponding to historical users in the target domain; the click record label is a click record of sample target domain content in the target domain; inputting the sample target domain user characteristics and the sample target domain content characteristics into a content recommendation model to be trained, and performing click prediction on sample target domain content; determining a click prediction loss value according to a click prediction result and the click record label; and adjusting parameters of the content recommendation model according to the click prediction loss value and continuing training until the training stopping condition is met.

A computer device comprising a memory and a processor, the memory storing a computer program, the processor implementing the following steps when executing the computer program:

and pushing the content to be recommended to the new user of the target domain.

A computer-readable storage medium, on which a computer program is stored which, when executed by a processor, carries out the steps of:

and pushing the content to be recommended to the new user of the target domain.

According to the content recommendation method, device, computer equipment and storage medium based on artificial intelligence, the source domain user characteristics corresponding to the new user of the target domain in the source domain are obtained; because the new user does not have perfect user characteristic information in the target domain, and the source domain user characteristics are obtained by performing characteristic extraction based on the interactive behavior information of the user identification in the source domain, the source domain user characteristics are subjected to characteristic mapping through a trained task-oriented mapping model, and thus the mapping user characteristics mapped from the source domain to the target domain can be effectively obtained. The server further carries out click prediction on the recommended content according to the mapping user characteristics and the target domain content characteristics of each candidate content in the target domain through the pre-trained content recommendation model, so that the click prediction on the candidate content in the target domain can be accurately and effectively carried out, and a more accurate and reasonable click prediction result can be obtained. The server screens the content to be recommended from the candidate content according to the click prediction result and pushes the content to be recommended to the new user in the target domain, so that when content recommendation is performed on the new user in the target domain, the mapping user characteristics of the new user in the target domain can be learned according to the source domain user characteristics of the new user in the source domain, content recommendation is performed by using the mapping user characteristics, the problem that the user characteristics of the new user in the target domain are insufficient can be effectively solved, and the content recommendation accuracy of the new user in the target domain can be effectively improved.

Drawings

FIG. 1 is a diagram of an embodiment of an application environment for an artificial intelligence based content recommendation method.

FIG. 2 is a flow diagram of a method for artificial intelligence based content recommendation in one embodiment.

Fig. 3 is a flowchart illustrating a method for artificial intelligence based content recommendation according to another embodiment.

FIG. 4 is a schematic flow chart of the first stage training step in one embodiment.

FIG. 5 is a flow diagram that illustrates a method for artificial intelligence based content processing, according to an embodiment.

FIG. 6 is a flow diagram illustrating a method for artificial intelligence based content processing in an exemplary embodiment.

FIG. 7 is a block diagram of an artificial intelligence based content recommendation device in one embodiment.

Fig. 8 is a block diagram showing a configuration of an artificial intelligence-based content recommendation apparatus according to another embodiment.

FIG. 9 is a block diagram of an artificial intelligence based content processing apparatus in one embodiment.

Fig. 10 is a block diagram showing the structure of an artificial intelligence-based content processing apparatus according to another embodiment.

FIG. 11 is a diagram illustrating an internal structure 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 is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.

The content recommendation method based on artificial intelligence can be applied to computer equipment. The computer device may be a terminal or a server. It can be understood that the content recommendation method based on artificial intelligence provided by the application can be applied to a terminal, can also be applied to a server, can also be applied to a system comprising the terminal and the server, and is realized through the interaction of the terminal and the server.

The artificial intelligence based content recommendation method can be applied to the application environment shown in fig. 1. Wherein the terminal 102 communicates with the server 104 via a network. The terminal 102 may be, but is not limited to, a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, and the like. The server 104 may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing basic cloud computing services such as a cloud service, a cloud database, cloud computing, a cloud function, cloud storage, a network service, cloud communication, middleware service, a domain name service, a security service, a CDN, a big data and artificial intelligence platform, and the like. The terminal 102 and the server 104 may be directly or indirectly connected through wired or wireless communication, and the application is not limited thereto.

Cloud computing (cloud computing) is a computing model that distributes computing tasks over a pool of resources formed by a large number of computers, enabling various application systems to obtain computing power, storage space, and information services as needed. The network that provides the resources is referred to as the "cloud". As a basic capability provider of cloud computing, a cloud computing resource pool (cloud platform, generally referred to as IaaS (Infrastructure as a Service) platform is established, the method mainly comprises the following steps of deploying various types of virtual resources in a resource pool for external customers to select and use, wherein the cloud computing resource pool mainly comprises the following steps: computing devices (which are virtualized machines, including operating systems), storage devices, network devices, are divided in logical functions, a Platform as a Service (Platform as a Service) layer can be deployed on an Infrastructure as a Service (IaaS) layer, a Software as a Service (SaaS) layer can be deployed on the PaaS layer, or the SaaS can be directly deployed on the IaaS layer, the PaaS is a Platform for Software operation, SaaS is a wide variety of business software, such as web portals, content recommendation systems, and the like.

Specifically, a user may send an access request to the server 104 through the corresponding terminal 102, and if the access request is initiated by a new user in the target domain, the server 104 obtains a source domain user feature corresponding to the new user in the target domain in the source domain, and performs feature mapping on the source domain user feature through a trained task-oriented mapping model to obtain a mapping user feature mapped from the source domain to the target domain; then, click prediction of recommended contents is carried out according to the mapping user characteristics and the target domain content characteristics of each candidate content in the target domain through a pre-trained content recommendation model; the server 104 filters the content to be recommended from the candidate content according to the result of the click prediction, and pushes the content to be recommended to the terminal 102 corresponding to the new user in the target domain.

It can be understood that, in the content recommendation method based on artificial intelligence in the embodiments of the present application, the task-oriented mapping model and the content recommendation model are trained by using a machine learning technique in an artificial intelligence technique, so that content recommendation can be accurately performed to a user in a target domain. Artificial Intelligence (AI) is a theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human Intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a manner similar to human intelligence. Artificial intelligence is the research of the design principle and the realization method of various intelligent machines, so that the machines have the functions of perception, reasoning and decision making.

Machine Learning (ML) is a multi-domain cross discipline, and relates to a plurality of disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and the like. The special research on how a computer simulates or realizes the learning behavior of human beings so as to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve the performance of the computer. Machine learning is the core of artificial intelligence, is the fundamental approach for computers to have intelligence, and is applied to all fields of artificial intelligence. Machine learning and deep learning generally include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and formal education learning. It can be understood that in some embodiments of the present application, the pre-trained task-oriented mapping model and the content recommendation model are trained by using a machine learning technique, and the task-oriented mapping model and the content recommendation model are trained based on the machine learning technique, so that a task-oriented mapping model with more accurate feature mapping and a content recommendation model with higher information recommendation accuracy can be trained.

In one embodiment, as shown in fig. 2, there is provided an artificial intelligence based content recommendation method, which is illustrated by applying the method to the server in fig. 1, and includes the following steps:

s202, obtaining source domain user characteristics corresponding to the new user of the target domain in the source domain.

The target domain, also referred to as a target domain, refers to a scene domain that needs to process a current target task. The source domain, also referred to as source realm, refers to similar information having an association with the target domain. In transfer learning, the domain of known knowledge is called the source domain and the domain of new knowledge to be learned is called the target domain. The goal of the transfer learning is to extract useful knowledge information from one or more source domain tasks and use the information on a new target domain task, and the transfer learning is particularly suitable for cross-domain recommendation, wherein the cross-domain recommendation is to combine a plurality of domain data to jointly act on the target domain recommendation. For example, the recommended task processing in the target domain is assisted by using the interaction data in an auxiliary domain, i.e. the source domain, so as to improve the recommendation effect.

It is to be understood that the domain concept can be represented in any distinguishable scenario, such as a recommendation application scenario for various recommended contents. The recommended content, that is, the recommended object, may include an article, resource promotion information, and the like. The articles may include virtual articles and real articles. In one embodiment, the item comprises an object capable of being traded by way of resource transfer. The resource popularization information is information for popularizing various resources, such as various recommendation information of articles, news content, video sharing content, webpage content, advertisement content, and the like.

There is usually a separate recommendation system for each product domain. For example, user-item interaction information exists in the network, and interaction between users is called social information, and the two types of interaction are heterogeneous, so that item recommendation from an information field to a social field user can be completed through a cross-field content recommendation process.

It can be understood that the new user of the target domain represents a user who does not have any interactive data in the service recommendation system corresponding to the current target domain, and the user has historical interactive data in the source domain.

Wherein, the feature, i.e. the feature vector, generally refers to the feature vector of the linear transformation. A linear transformation can be fully described by its eigenvalues and eigenvectors, a set of eigenvectors of the same eigenvalue being referred to as the eigenspace. Each domain has a respective feature space, for example, feature vector representations of the same user in different recommendation systems are different, that is, feature spaces are different.

It should be noted that the source domain user feature refers to a user feature of a new user of the target domain in the source domain, and the source domain user feature is obtained by performing feature extraction on historical interaction behavior information of the user in the source domain. The source domain user characteristics comprise a plurality of attribute characteristics of the user, for example, characteristics corresponding to information including gender, age, frequent residence, interest preference, behavior type and the like.

In one embodiment, the source domain user characteristics may further include static user characteristics and dynamic user characteristics corresponding to the user in the source domain, where the static user characteristics refer to identity attribute characteristics of the user, and include attribute characteristics corresponding to information such as gender, age, frequent residence, and the like; the dynamic user characteristics refer to the behavior characteristics of the user, and comprise characteristics of behavior type, behavior preference, item preference and the like. Wherein the dynamic user characteristics may reflect the interests and preferences of the user.

Specifically, an application program corresponding to the target domain is run in the terminal of the user, and the user can initiate an access request to the service recommendation system corresponding to the target domain through the application program in the terminal. The server may further perform new user identification on the user corresponding to the access request, and specifically, may query whether the user has historical interaction data in the target domain, thereby determining whether the user is a new user. And the server corresponding to the target domain receives the access request sent by the terminal, and when the user corresponding to the access request is identified to be a new user in the target domain, the user characteristics of the source domain corresponding to the new user in the source domain with the historical interactive data are obtained.

In one embodiment, obtaining source domain user characteristics corresponding to a new user of a target domain in a source domain includes: according to the user identification of the new user in the target domain, searching a source domain with interactive behavior information corresponding to the user identification in a plurality of source domains; acquiring source domain user characteristics corresponding to user identification from a source domain; the source domain user characteristics are obtained by performing characteristic extraction based on the interaction behavior information of the user identification in the source domain.

The user identifier refers to an account identifier used by the user in the application system to identify the user identity, and may be, for example, an account name, a mobile phone number, an application number, and other identifiers. The user identification is typically a unique identifier that identifies the identity of the user. The user identities of the same user in different domain applications may be the same.

When the server identifies that the user who initiates the access request in the target domain is a new user, the new user may have historical interactive behavior information in applications in other domains. And the server searches a source domain with interactive behavior information corresponding to the user identification in a plurality of source domains according to the user identification of the new user, and further acquires the source domain user characteristics corresponding to the user identification from the source domain.

The source domain user characteristics in the source domain are obtained by performing characteristic extraction according to the interactive behavior information of the user identification in the source domain. The source domain comprises the source domain user characteristics corresponding to the extracted user identification, so that the server can directly acquire the corresponding source domain characteristics from the source domain according to the user identification.

And S204, performing feature mapping on the source domain user features through the trained task-oriented mapping model to obtain mapping user features mapped from the source domain to the target domain.

The task-oriented mapping model is a machine learning model which is trained and has the capability of carrying out task-oriented feature mapping on the source domain user features so as to map the source domain user features to the user features corresponding to the target domain. The task-oriented mapping model is obtained by training historical behavior information of historical users shared in the source domain and the target domain. Specifically, the method comprises the steps that a sample source domain user characteristic of a historical user in a source domain and a sample target domain content characteristic in a target domain are used as training samples, a click record corresponding to the sample target domain content in the target domain is used as a training label, and a task guide for performing click prediction on the sample target domain content is used as a target to obtain a task guide mapping model through training.

The task guide mapping model comprises a mapping function between a source domain and a target domain, and the mapping function is obtained by predicting corresponding task guide based on clicking of sample target domain content and performing transfer learning.

It is understood that feature mapping refers to mapping data of a source domain and a target domain from an original feature space into a new feature space, that is, mapping a feature space in the source domain into a feature space in the target domain, where the source domain data and the target domain data have the same distribution. Therefore, model training can be performed by better utilizing the existing labeled data samples in the source field in a new space. As long as a certain incidence relation exists between the source domain and the target domain, when the task guide mapping model in the target domain is trained, the reuse and migration of the learned knowledge among similar or related domains can be realized by means of the knowledge extracted from the data and the characteristics of the source domain.

And after obtaining the source domain user characteristics corresponding to the source domain of the new user, the server inputs the obtained source domain user characteristics to a pre-trained task guide mapping model, and performs characteristic mapping on the source domain user characteristics through the task guide mapping model, so that the mapping user characteristics mapped from the source domain to the target domain can be obtained.

The method comprises the steps of performing feature mapping processing on source domain user features in a source domain through a pre-trained task-oriented mapping model, and mapping relevant user features of a new user in the source domain into a feature space in a target domain, so that the mapped user features obtained through mapping have relevant user features in the target domain, and therefore the user features of the new user aiming at the target domain content in the target domain can be obtained, the fact that the new user does not have relevant user features in the target domain is effectively made up, and accuracy of content recommendation of the new user in the target domain can be effectively improved.

And S206, performing click prediction on recommended contents according to the mapping user characteristics and the target domain content characteristics of each candidate content in the target domain through the pre-trained content recommendation model.

It is understood that a large number of candidate contents to be recommended, i.e., target domain contents, are included in the target domain. The Content may include, but is not limited to, an article, a picture, a video, an advertisement, a web page, an application program or UGC (User Generated Content), that is, User Generated Content, and specifically, the Content to be recommended may include information Content in various forms, such as text, an image, audio, a video, a web page, and in various combinations of forms.

It can be understood that the content recommendation model is a machine learning model with a click prediction capability for the content to be recommended in the target domain, which is trained in advance. The content recommendation model can be obtained by training historical behavior information of historical users in the target domain. The content recommendation model is used for recommending the content to be recommended in the target domain to the user according to the user characteristics of the user in the target domain.

The server obtains the mapping user characteristics mapped from the source domain to the target domain through the task-oriented mapping model, and therefore the user characteristics of the new user in the target domain are obtained. And the server further carries out click prediction on the recommended content according to the mapping user characteristics and the target domain content characteristics of each candidate content in the target domain through a pre-trained content recommendation model.

And respectively calculating the click prediction value corresponding to each candidate content in the target domain by the new user through a content recommendation model by the server, thereby obtaining the click prediction result corresponding to each candidate content. Therefore, the interest degree and the probability of possible click of the new user on each candidate content can be accurately predicted, corresponding content is recommended to the new user according to the prediction result, and the content recommendation accuracy can be effectively improved.

And S208, screening the content to be recommended from the candidate content according to the click prediction result.

The click prediction result can reflect the predicted interest degree of the new user for each candidate content, and the click rate prediction is used for predicting the click condition of the target domain content and judging the probability of clicking the content by the user. The click through rate prediction value is used for recommending the content in the target domain.

And after the server obtains the click prediction result corresponding to each candidate content in the target domain, screening the content to be recommended from the candidate contents according to the click prediction result. Specifically, the server screens out candidate contents with a click prediction value reaching a preset threshold value, and takes the screened candidate contents with the click prediction value reaching the preset threshold value as contents to be recommended for the new user.

S210, pushing the content to be recommended to the new user of the target domain.

And after the server screens out the content to be recommended, pushing the content to be recommended to the new user of the target domain. Further, the server may also determine a recommendation order of the content to be recommended according to a result of the click prediction, and specifically may perform descending order on each candidate content according to a click prediction value corresponding to each candidate content, so as to push the content to be recommended to the new user in the target domain according to the ordered recommendation order. And after the terminal corresponding to the new user obtains the contents to be recommended pushed by the server, displaying the corresponding recommended contents on the interface of the terminal in sequence according to the recommendation sequence.

In traditional transfer learning, a mapping relationship between a user vector and an article vector between two domains is generally learned respectively, the user vector is also a user feature, and the article vector is also an article feature. Namely, the mapping relation between the source domain user characteristics and the target domain user characteristics and the mapping relation between the source domain article characteristics and the target domain article characteristics are respectively learned to learn the mapping functions between the two domains, and the method is based on the loss function learning mapping function of mapping guidance. However, the target domain user vector of the target domain may not be accurate enough to reflect the user interest well, so that the mapping feature mapped from the source domain to the target domain may have noise influence, and the user feature obtained by using the mapping may have a large error.

In the task guide mapping model in this embodiment, a mapping relationship between a user vector and an article vector in two fields is not learned, but a historical user shared in a source domain and a target domain is used, a sample source domain user characteristic in the source domain and a sample target domain content characteristic in the target domain are used as training samples, a click record corresponding to a sample target domain content in the target domain is used as a training label, a corresponding task guide is predicted based on a click on the sample target domain content, and training is performed by using a loss of the task guide. Therefore, the target domain user characteristics in the target domain are avoided, and the content recommendation accuracy of the new user in the target domain can be effectively improved.

In the content recommendation method based on artificial intelligence, source domain user characteristics corresponding to a new user of a target domain in a source domain are obtained; because the new user does not have perfect user characteristic information in the target domain, and the source domain user characteristics are obtained by performing characteristic extraction based on the interactive behavior information of the user identification in the source domain, the source domain user characteristics are subjected to characteristic mapping through a trained task-oriented mapping model, and thus the mapping user characteristics mapped from the source domain to the target domain can be effectively obtained. The server further carries out click prediction on the recommended content according to the mapping user characteristics and the target domain content characteristics of each candidate content in the target domain through the pre-trained content recommendation model, so that the click prediction on the candidate content in the target domain can be accurately and effectively carried out, and a more accurate and reasonable click prediction result can be obtained. The server screens the content to be recommended from the candidate content according to the click prediction result and pushes the content to be recommended to the new user in the target domain, so that when content recommendation is performed on the new user in the target domain, the mapping user characteristics of the new user in the target domain can be learned according to the source domain user characteristics of the new user in the source domain, content recommendation is performed by using the mapping user characteristics, the problem that the user characteristics of the new user in the target domain are insufficient can be effectively solved, the content recommendation accuracy of the new user in the target domain can be effectively improved, and the recommendation effect of cross-domain cold start recommendation is effectively improved.

In one embodiment, performing feature mapping on source domain user features through a trained task-oriented mapping model to obtain mapping user features mapped from a source domain to a target domain, includes: inputting the source domain user characteristics into a trained task guide mapping model, and performing characteristic matrix transformation on the source domain user characteristics through a mapping layer of the task guide mapping model; and outputting the mapping user characteristics for mapping the source domain user characteristics to the target domain according to the result of the characteristic matrix transformation.

It will be appreciated that the feature matrix transforms, representing a mapping from one spatial vector to another and corresponding vector calculations are performed. The feature matrix transformation may be a non-linear transformation to map user features originating in the source domain to the feature space in the destination domain, thereby transforming the mapped user features in the feature space of the destination domain.

The task-oriented mapping model may include a mapping layer, and the mapping layer includes a linear layer, an activation function, and a linear layer, wherein one linear layer, the activation function, and the linear layer may be regarded as a matrix transformation layer, and the mapping layer may further include a plurality of matrix transformation layers.

After the server inputs the source domain user characteristics to the trained task guide mapping model, the characteristic matrix transformation is carried out on the source domain user characteristics through a linear matrix in a mapping layer of the task guide mapping model, and after the linear matrix transformation is carried out, the original source domain user characteristics can be transformed into vector representation in a characteristic space of a target domain, so that the mapping user characteristics for mapping the source domain user characteristics to the target domain are obtained. The pre-trained task-oriented mapping model comprises user feature distribution and corresponding mapping relation between the source domain and the target domain corresponding to the source domain content and the target domain content. For example, the linear matrix transformation process may specifically adopt a method based on eigen mapping of a regenerative kernel hilbert space or a method based on laplacian eigen mapping.

In the embodiment, the source domain user characteristics are subjected to feature mapping processing through the pre-trained task-oriented mapping model, so that the user characteristics of the new user in the target domain for the target domain content can be obtained, the problem that the user characteristics of the new user in the target domain are insufficient can be effectively solved, and the accuracy of content recommendation of the new user in the target domain can be effectively improved.

In one embodiment, the performing, by using a pre-trained content recommendation model, click prediction of recommended content according to the mapping user characteristics and the target domain content characteristics of each candidate content in the target domain respectively includes: inputting the mapping user characteristics and the target domain content characteristics of the candidate content in the target domain into a pre-trained content recommendation model, and predicting the interest degree of a new user for each candidate content through the content recommendation model; and carrying out click prediction scoring on each candidate content according to the interestingness to obtain a click prediction result corresponding to each candidate content.

The pre-trained content recommendation model may be a model formed by an artificial neural network. The Neural Network model includes, for example, a CNN (Convolutional Neural Network) model, a DNN (Deep Neural Network) model, an RNN (Recurrent Neural Network) model, an LSTM (Long Short-Term Memory Neural Network) model, and the like. The content recommendation model may also be a combination of neural network models.

It is understood that a deep neural network typically includes an input layer, an implicit layer, and an output layer, with a fully connected relationship between layers. A prediction layer is also included in the content recommendation model. The prediction Layer may include a multi-head self-attention mechanism Layer (multihead self-attention-mechanism), a position-wise full link Layer (position-wise fed Forward), and the middle Layer may include a multi-Layer residual error Normalization Layer (Layer Normalization) and an FFN (Feed Forward neural network), and each sub-Layer has a residual error connection for Layer Normalization. The prediction layer comprises a prediction function, the prediction function can be a preset loss function, and the loss of the content recommendation model can be learned to obtain corresponding loss by adopting a mode of sequencing learning based on Point-wise points, Pair-wise synchronous learning based on Pair-wise points, List-wise List learning based on List-wise or Metric learning based on Metric-loss. The loss function may specifically adopt a Sigmoid function, a Tanh function, a ReLu function, a Softmax function, or a cross entropy loss function. The prediction layer is used for predicting the click probability of the user on the candidate content in the target domain.

After obtaining the mapping user characteristics corresponding to the new user, the server inputs the mapping user characteristics and the target domain content characteristics corresponding to a plurality of candidate contents in the target domain into the content recommendation model, predicts the interest degree of the user for each candidate content in the target domain according to the mapping user characteristics of the user through each neural network of a prediction layer in the content recommendation model, predicts the click probability of each candidate content according to the predicted interest degree, obtains the click prediction score of each candidate content, and the click prediction score is the click prediction result of the candidate content.

For example, each User includes a corresponding User identification, e.g., denoted User ID, and each candidate content, i.e., target domain content, also includes a corresponding content identification, e.g., denoted Item ID. The source domain User characteristics of the User in the source domain may be represented as User ID embedding, and the target domain content characteristics of the candidate content in the target domain may be represented as Item IID embedding. And carrying out feature Mapping on the User ID embedding through the trained task-oriented Mapping model, and Mapping the User ID embedding from the source domain Mapping to the feature space of the target domain to obtain the mapped Mapping User features, wherein the Mapping User features can be expressed as Transformed ID embedding. And then inputting the transformated ID embedding and the plurality of Item IID embedding into a pre-trained content recommendation model in a target domain, carrying out click prediction on each Item IID embedding, and screening out the content to be recommended for a new user according to the result of the click prediction.

In the embodiment, the content recommendation model is used for performing click prediction on the recommended content for the new user in the target domain according to the mapping user characteristics and the target domain content characteristics corresponding to the candidate content, so that the interest degree and the probability of possible click of the new user for each candidate content can be accurately predicted, the corresponding content is recommended to the new user according to the prediction result, and the content recommendation accuracy can be effectively improved.

In one embodiment, as shown in fig. 3, another artificial intelligence based content recommendation method is further provided, comprising the steps of:

s302, acquiring an access request corresponding to a new user in the first sub-application, wherein the access request comprises a user identifier; the first sub-application corresponds to a target domain; the first sub-application runs in an environment provided by the parent application.

S304, searching a second sub-application running in the parent application and having the interactive behavior information corresponding to the user identifier according to the user identifier; the second sub-application corresponds to the source domain; the second sub-application comprises source domain user characteristics obtained based on the interactive behavior information of the user identification.

S306, obtaining the source domain user characteristics corresponding to the new user of the target domain in the source domain.

And S308, performing feature mapping on the source domain user features through the trained task-oriented mapping model to obtain mapping user features mapped from the source domain to the target domain.

S310, through a pre-trained content recommendation model, click prediction of recommended content is conducted according to the mapping user characteristics and the target domain content characteristics of each candidate content in the target domain.

And S312, screening the content to be recommended from the candidate content according to the click prediction result.

S314, pushing the content to be recommended to the new user of the target domain.

The parent application refers to an application that can run independently, and is a native application program that runs directly on the operating system. The parent application may specifically be an application program that carries the child application, and may provide an execution environment for the application executed by the child application. The sub application refers to an application that can be used without downloading and installing, but the sub application needs to be run on the parent application.

It is understood that the sub-application can be various business application scenes attached to the parent application, and can also be a sub-application running in the parent application. Each sub-application corresponds to a different domain. The parent application and the child application include, but are not limited to, an instant messaging application, an SNS application, a short video application, a long video application, a game application, an article application, a music sharing application, and a UGC application.

Taking a parent application as an instant messaging application as an example, the parent application includes a plurality of child applications, for example, the child applications may be an article application, a short video application, a game application, and the like.

It can be understood that the user has a corresponding user identifier in the parent application, and the user identifier of the user in each child application of the parent application may be the same as the user identifier in the parent application. For example, may be the same application account identification, such as an application account; or an identity such as a cell phone number.

The terminal corresponding to the user runs the parent application, the user can send an access request to the first child application in the parent application through the corresponding terminal, and the access request is processed through the first child application, so that candidate content in the first child application is recommended to the user.

And after acquiring the access request corresponding to the new user in the first sub-application, the server identifies whether the user is the new user according to the user identification carried in the access request. And if the user is a new user, the server searches a second sub-application which runs in the parent application and has the interactive behavior information corresponding to the user identifier according to the user identifier, takes the first sub-application as a target domain and takes the second sub-application as a source domain. And the second sub-application comprises source domain user characteristics obtained by performing characteristic extraction based on the interactive behavior information of the user identification.

And then, the server acquires the source domain user characteristics corresponding to the new user of the target domain in the source domain from the second sub-application. The server further performs feature mapping on the source domain user features through the trained task-oriented mapping model to obtain mapping user features which are mapped from the second sub-application to the first sub-application. And performing click prediction on recommended contents according to the mapping user characteristics and the target domain content characteristics of each candidate content in the first sub-application through a pre-trained content recommendation model. And the server further screens the content to be recommended from the candidate content according to the click prediction result and pushes the screened content to be recommended in the first sub-application to the new user.

In this embodiment, for a new user in a first sub-application running in a parent application, the first sub-application is used as a target domain, and according to a user identifier of the new user, a source domain user feature is obtained from a second sub-application corresponding to interaction behavior information of the new user in the parent application, so that a user feature that the new user has interaction behaviors in other fields can be effectively obtained, a more accurate mapping user feature of the new user in the target domain is obtained by performing feature mapping on the source domain user feature in the second sub-application, and then content in the first sub-application can be recommended to the new user according to the mapping user feature, so that the recommendation accuracy of the new user is effectively improved.

In one embodiment, the content recommendation model is obtained by a first stage training step, as shown in fig. 4, which is a schematic flow chart of the first stage training step in one embodiment, and includes the following steps:

s402, acquiring historical interaction information and corresponding click record labels in a target domain; the historical interactive information comprises sample target domain user characteristics and sample target domain content characteristics corresponding to historical users in the target domain; the click record label is a click record of sample target domain content in the target domain.

S404, inputting the user characteristics of the sample target domain and the content characteristics of the sample target domain into a content recommendation model to be trained, and performing click prediction on the content of the sample target domain.

And S406, determining a click prediction loss value according to the click prediction result and the click record label.

And S408, adjusting parameters of the content recommendation model according to the click prediction loss value and continuing training until the training stopping condition is met.

It can be understood that the first-stage training refers to one of the stages of the model training process, the first-stage training is used to distinguish the second-stage training and the sequence of the first-stage training and the second-stage training, that is, the first-stage training is required first, and then the second-stage training is required, that is, the content recommendation model is required to be trained first, and then the task oriented mapping model is required to be trained.

It can be understood that the historical interaction information refers to information of user information behaviors of a historical user and a corresponding service system in an application corresponding to a target domain, and the user information behavior refers to behaviors of demand expression, information acquisition, information utilization and the like expressed when the user seeks required information. The historical interaction information is historical interaction data generated in the target domain application by the user in the past period.

The historical interaction information in the target domain comprises user identifications corresponding to a plurality of historical users and historical interaction contents corresponding to the user identifications of the historical users, and each historical user can comprise the historical interaction information corresponding to a plurality of historical interaction contents. The historical interaction information can be obtained from a preset sample library, and also can be obtained from application platforms of various target domains, such as historical interaction information of historical users in applications of social networks, video sharing networks, community forums, blogs and the like. The historical interactive content may include various forms of content such as text, articles, pictures, audio, video, advertisements, and so forth.

After obtaining the historical interaction information and the corresponding click record label in the target domain, the server takes the historical interaction information as a training sample of a training content recommendation model, takes the click record label as a training label, and takes the historical interaction content as sample target domain content. The historical interaction information comprises sample target domain user characteristics and sample target domain content characteristics corresponding to historical users in the target domain, and the click record label is a click record of the content of each sample target domain in the historical interaction information.

It can be understood that the sample target domain user features and the sample target domain content features may be obtained by pre-extraction, or the historical interaction information may be directly input into the content recommendation model, and the sample target domain user features and the sample target domain content features are respectively extracted by performing feature extraction on the historical interaction information through the content recommendation model.

The server inputs the obtained sample target domain user characteristics and the sample target domain content characteristics into a content recommendation model, the click prediction result is obtained by predicting the click probability of the historical user on the sample target domain content according to the sample target domain user characteristics, and the click prediction loss value is determined according to the click prediction result and the click record label. And the server further adjusts parameters of the content recommendation model according to the click prediction loss value and continues training until the training stopping condition is met, and then the training is finished.

The result of the click prediction includes a click prediction value of each sample target domain content, and specifically, a click prediction loss value may be calculated according to a difference between the click prediction value of each sample target domain content and the click record label. The click prediction loss value can be measured by a loss function, for example, a mean absolute value loss function, a smooth mean absolute error, a cross entropy loss function, and the like can be selected as the loss function.

It is understood that the training condition is a condition for ending the model training. The training stopping condition may be that a preset number of iterations is reached, or that the predicted performance index of the content recommendation model after the parameters are adjusted reaches a preset index, so as to implement accurate recommendation of the content in the target domain.

In the embodiment, the content recommendation model is trained through the sample target domain user characteristics and the sample target domain content characteristics and the click record label, and parameters in the content recommendation model are gradually adjusted according to the click prediction loss value. Therefore, in the parameter adjustment process, the content recommendation model can capture the implicit relation between the user characteristics and the content characteristics and the click rate, the click prediction accuracy of the content is effectively improved, and the content recommendation model with high recommendation accuracy can be trained.

In one embodiment, the server may obtain an access request sent by a user to the target domain through the terminal, and when the server identifies that the user corresponding to the access request is not a new user, the server may obtain the interactive behavior information indicating that the user has a history in the target domain. The server can directly acquire the interactive behavior information corresponding to the user to perform feature extraction, so as to obtain the corresponding target domain user features. And the server predicts the interest degree of the user for each candidate content according to the target domain user characteristics of the user and the target domain content characteristics corresponding to the candidate content in the target domain through the trained content recommendation model, and outputs the click prediction scores of each candidate content. And the server further screens out the contents to be recommended which meet the pushing conditions from the candidate contents according to the predicted click prediction scores, and pushes the screened contents to be recommended to the terminal corresponding to the user.

In one embodiment, the task-oriented mapping model is obtained by a second-stage training step, the second-stage training step including: acquiring historical interaction information and a corresponding click record label; the historical interaction information comprises sample source domain user characteristics of a historical user in a source domain and sample target domain content characteristics in a target domain; performing feature mapping on the sample source domain user features through a task-oriented mapping model to be trained to obtain sample mapping user features mapped from a source domain to a target domain; inputting the sample mapping user characteristics and the sample target domain content characteristics into a pre-trained content recommendation model, and performing click prediction on the sample target domain content; determining a task guide loss value according to a result of the click prediction and the click record label; and adjusting parameters of the task guide mapping model according to the task guide loss value and continuing training until the training stopping condition is met.

Specifically, the server obtains sample source domain user characteristics of the historical users in the source domain and sample target domain content characteristics in the target domain from historical behavior information of the historical users shared in the source domain and the target domain, and click record labels corresponding to the sample target domain contents.

And then, inputting the sample source domain user characteristics into a to-be-trained task guide mapping model, and performing characteristic mapping on the sample source domain user characteristics through the to-be-trained task guide mapping model to map the relevant user characteristics of each historical user in the source domain into a characteristic space in the target domain, so as to obtain the sample mapping user characteristics mapped from the source domain to the target domain.

The server obtains sample mapping user characteristics mapped from a source domain to a target domain through a task-oriented mapping model, then inputs the sample mapping user characteristics and the sample target domain content characteristics into a pre-trained content recommendation model, and respectively calculates click prediction values of each historical user corresponding to each sample target domain content in the target domain through the content recommendation model, so as to obtain click prediction results of each historical user corresponding to each sample target domain content.

And the server further determines a task guide loss value according to the result of the click prediction and the click record label, adjusts parameters of the task guide mapping model according to the task guide loss value and continues training so as to carry out iterative training on the task guide mapping model until the training stopping condition is met, and then the training is finished.

The loss of the task guide mapping model can be consistent with the loss of the content recommendation model trained in the first stage, and is based on the loss of the task guide corresponding to the click prediction of the sample target domain content.

It is understood that the training condition is a condition for ending the model training. By avoiding the training of the task-oriented mapping model by the target domain user features in the target domain, the error of mapping the source domain user features to the mapping user features in the target domain is reduced, so that the accuracy of the task-oriented mapping model for mapping the source domain user features to the user features in the target domain can be effectively improved.

In one embodiment, as shown in fig. 5, there is provided an artificial intelligence based content processing method, which is described by taking the method as an example applied to the server in fig. 1, and includes the following steps:

s502, acquiring historical interaction information and a corresponding click record label; the historical interaction information comprises sample source domain user characteristics of the historical user in the source domain and sample target domain content characteristics in the target domain.

The historical interaction information comprises historical behavior information of historical users shared in a source domain and a target domain, sample source domain content and sample target domain content, and click record labels corresponding to the sample source domain content and the sample target domain content respectively and respectively for the historical users.

In one embodiment, obtaining historical interaction information and corresponding click record tags includes: acquiring user identifications corresponding to historical users shared in a source domain and a target domain; respectively acquiring historical interaction information of historical users in a source domain and a target domain according to the user identification; extracting sample source domain user characteristics of a historical user in a source domain, sample target domain content characteristics of the historical user in a target domain and click record labels corresponding to the historical interaction information in the target domain from the historical interaction information; and the click record label is a click record corresponding to each sample target domain content in the historical interaction information of the target domain.

The server may specifically search user identifiers corresponding to historical users shared in the source domain and the target domain, and then obtain historical interaction information in the source domain and the target domain according to the user identifiers, respectively. The server further extracts the characteristics of the historical interactive information, extracts the sample source domain user characteristics of each historical user in the source domain and the sample target domain content characteristics in the target domain.

In the process of training the task guide mapping model, the server takes the sample source domain user characteristics of each historical user in the source domain and the sample target domain content characteristics in the target domain as training samples of the training task guide mapping model, and takes the click record labels corresponding to the sample target domain contents as training labels.

For example, there are N history users shared by the two domains, and each history user has M interactions in the target domain on average, that is, there are interactions on the target domains of M samples, and then the samples corresponding to the target domains of N × M samples can be directly used as training samples to train the task-oriented mapping model. Therefore, the number of training samples is greatly increased, the problem of overfitting can be effectively solved, and the training accuracy of the task guide mapping model can be effectively improved.

S504, performing feature mapping on the sample source domain user features through a task-oriented mapping model to be trained to obtain sample mapping user features mapped from a source domain to a target domain.

The server acquires sample source domain user characteristics and sample target domain content characteristics corresponding to each historical user, inputs the sample source domain user characteristics into a task guide mapping model to be trained, and performs characteristic mapping on the sample source domain user characteristics through the task guide mapping model to be trained so as to map the relevant user characteristics of each historical user in the source domain into a characteristic space in the target domain, thereby obtaining the sample mapping user characteristics mapped from the source domain to the target domain.

S506, inputting the sample mapping user characteristics and the sample target domain content characteristics into a pre-trained content recommendation model, and performing click prediction on the sample target domain content.

And S508, determining a task guide loss value according to the click prediction result and the click record label.

And S510, adjusting parameters of the task guide mapping model according to the task guide loss value and continuing training until the training stopping condition is met.

The server obtains sample mapping user characteristics mapped from a source domain to a target domain through a task-oriented mapping model, then inputs the sample mapping user characteristics and the sample target domain content characteristics to a pre-trained content recommendation model, and respectively calculates click prediction values of each historical user corresponding to the content of each sample target domain in the target domain according to the sample mapping user characteristics and the target domain content characteristics of the sample target domain content in each target domain through the content recommendation model, so that click prediction results of each historical user corresponding to the content of each sample target domain are obtained.

And the server further determines a task guide loss value according to the click prediction result and the click record label, and specifically calculates the task guide loss value according to the difference between the click prediction value of each sample target domain content and the corresponding click probability label. The click prediction loss value may be measured by using a target loss function, for example, a function such as an average absolute value loss function, a smoothed average absolute error, a cross entropy loss function, or the like may be selected as the loss function.

In one embodiment, the trained target loss function may be as follows:

wherein the content of the first and second substances,L _θis the objective loss function of the task-oriented mapping model,Dais a set of sample data that is,xis the time period for each sample in the sample data,L _taskis to predict the loss of task guidance for the task. For example, the sample data set includes N × M samples, where there are N history users shared by the two domains, and each history user has M interactions in the sample target domain of the target domain, then samples corresponding to the N × M sample target domain contents are extracted, where M is the number of users and N is the number of items corresponding to the sample target domain contents.

For example, withL _taskThe loss function is based on the loss learned in a Point-wise manner as an example,L _taskthe loss function may be as follows:

wherein, assuming that there are M users and N items corresponding to the sample target content, i and j are respectively used to correspondingly represent the index,u _ia hidden vector representing each of the users is represented,v _ia hidden vector representing the corresponding item of each sample target domain content,r _ijrepresenting the click prediction of user i on item j. Matrix R_M×NAnd the scoring matrix represents user-item interaction and comprises click record labels of all known users for the items corresponding to the contents of the target domains of all samples.R _ijAnd scoring the content of the jth sample target domain for the ith user. Wherein the content of the first and second substances,R _ija value of 1 indicates that a positive relationship exists between the user i and the item j;R _ija value of 0 represents that there is a negative relationship, or that the relationship is unknown, between user i and item j. σ is a function that converts the output into a probability, which may be, for example, a sigmoid function. Finally, calculating the dot product between the hidden vector of the user and the hidden vector of the content of the sample target domain, and predicting the result of clicking with the real valueR _ijAnd comparing and optimizing the network parameters in the task-oriented mapping model.

It is understood that in other embodiments, the corresponding loss may be learned by Pair-wise synchronous learning based on Pair-wise, List-wise, or Metric-loss Metric learning, or other functions such as Softmax function or cross entropy loss function may be used as the loss function, which is not limited herein.

And the server further adjusts the parameters of the task guide mapping model according to the task guide loss value and continues training so as to carry out iterative training on the task guide mapping model until the training stopping condition is met, and then the training is finished. It is understood that the training condition is a condition for ending the model training. The training stopping condition may be that a preset number of iterations is reached, or that the predicted performance index of the content recommendation model after the parameters are adjusted reaches a preset index, so as to implement accurate recommendation of the content in the target domain.

In one embodiment, determining a task oriented loss value based on the result of the click prediction and the click record label includes: determining a task guide loss value according to the difference between the click prediction result and the click record label;

adjusting parameters of the task-oriented mapping model according to the task-oriented loss value and continuing training, wherein the method comprises the following steps: and adjusting parameters of the task guide mapping model according to the task guide loss value so that the difference between the click prediction result and the click record label is reduced in the iterative training process of the task guide mapping model.

And in the process of training the task guide mapping model, the server calculates the task guide loss value according to the difference between the result of each click prediction and the click record label of the corresponding sample target domain content. And then updating the parameters of the task guide mapping model according to the contrast loss value.

Specifically, parameters of the task guide mapping model are updated according to the task guide loss value, so that in the process of performing iterative training on the task guide mapping model, the difference between the click prediction result and the click record label is continuously reduced, the sample mapping characteristics are closer and closer, and the target domain user characteristics corresponding to the sample target domain content of the click record label are provided. Therefore, errors of mapping user features for mapping the source domain user features to the target domain are reduced, and accuracy of mapping the user features to the target domain based on click prediction task guidance can be effectively improved.

It is to be understood that the task-oriented mapping model may be trained by using Supervised Learning (Semi-Supervised Learning), Semi-Supervised Learning (unsupervised Learning), open Learning (unsupervised Learning), and less Learning training, which is not limited herein.

In one embodiment, a supervised model training approach may be employed. Under the supervision type learning, data input into the model are training data, each set of training data has a definite identification, namely each set of samples comprises a sample source domain user characteristic, a sample target domain content characteristic and a click record label corresponding to the sample target domain content corresponding to a historical user. In the process of training the task-oriented mapping model by adopting a supervised learning mode, the prediction result of each time is compared with the click record label of the corresponding sample target domain content to obtain the corresponding difference, and the parameters of the task-oriented mapping model are continuously adjusted according to the difference until the prediction result of the task-oriented mapping model reaches an expected accuracy rate.

Taking a back propagation algorithm as an example, in the process of iteratively training the task-oriented mapping model, based on the back propagation algorithm, the parameters are updated towards the gradient descending direction, the weight and the bias are adjusted to minimize the overall error, and the parameters of the task-oriented mapping model are gradually adjusted to iteratively train the task-oriented mapping model.

In another embodiment, a semi-supervised model training approach may be employed. A large amount of unmarked data can be adopted in semi-supervised learning, and simultaneously, marked data are adopted for pattern recognition. By adopting semi-supervised learning, the resource consumption of data processing can be effectively reduced, and higher accuracy can be brought.

According to the content processing method based on artificial intelligence, the sample source domain user characteristics in the source domain and the sample target domain content characteristics in the target domain are used as training samples by using the history users shared in the source domain and the target domain, the click records corresponding to the sample target domain content in the target domain are used as training labels, the task guide corresponding to the click prediction of the sample target domain content is used for training by using the loss value of the task guide, so that the target domain user characteristics in the target domain are avoided, the error of mapping the source domain user characteristics to the mapping user characteristics in the target domain is reduced, the accuracy of mapping the user characteristics to the target domain based on the click prediction task guide can be effectively improved, and the accuracy of content recommendation for new users in the target domain can be effectively improved.

In one embodiment, performing feature mapping on sample source domain user features through a task-oriented mapping model to be trained to obtain sample mapping user features mapped from a source domain to a target domain, includes: inputting the sample source domain user characteristics into a task guide mapping model to be trained, and performing characteristic matrix transformation on the sample source domain user characteristics through a mapping layer of the task guide mapping model; and outputting sample mapping user characteristics for mapping the sample source domain user characteristics to the target domain according to the result of the characteristic matrix transformation.

The task-oriented mapping model comprises a mapping layer, wherein the mapping layer comprises a linear layer, an activation function and a linear layer, one linear layer, the activation function and the linear layer can be regarded as a matrix transformation layer, and the mapping layer can further comprise a plurality of matrix transformation layers.

And after obtaining the sample source domain user characteristics and the sample target domain content characteristics corresponding to each historical user, the server inputs the sample source domain user characteristics into the task guide mapping model to be trained, and performs characteristic mapping on the sample source domain user characteristics through the task guide mapping model to be trained.

Specifically, the sample source domain user characteristics are subjected to characteristic matrix transformation through a mapping layer of a task-oriented mapping model, so that the relevant user characteristics of each historical user in the source domain are mapped into a characteristic space in a target domain, and thus the sample mapping user characteristics mapped from the source domain to the target domain are obtained.

Specifically, the source domain includes a category and a corresponding weight of each source domain content, and the target domain also includes each target domain content and a corresponding weight. The source domain user characteristics include interest preferences of the user for the content of each source domain in the source domain, for example, the interest preferences may be weighted according to corresponding categories and weights based on the content of the target domain preferred by the user, so as to obtain the interest degree of the historical user for the content of each target domain. The pre-trained task-oriented mapping model comprises user feature distribution and corresponding mapping relation between the source domain and the target domain corresponding to the source domain content and the target domain content. The task guide is predicted by clicking the sample target domain content according to the source domain user characteristics, and the task guide mapping model is iteratively trained by using the loss of the task guide, so that the sample source domain characteristics of the user in the source domain can be obtained, and the mapping relation corresponding to the interest of the sample target domain content in the target domain is aimed at. Therefore, the accuracy of mapping to the user characteristics in the target domain based on click prediction task guidance can be effectively improved.

In one embodiment, inputting the sample mapping user characteristics and the sample target domain content characteristics into a pre-trained content recommendation model to perform click prediction of the sample target domain content, including: inputting the sample mapping user characteristics and the sample target domain content characteristics into a pre-trained content recommendation model, and predicting the sample interest degree of the historical user on the content of each sample target domain through the content recommendation model; and performing click prediction on the content of each sample target domain according to the interest degree of the sample to obtain a click prediction result corresponding to the content of each sample target domain.

After the server obtains sample mapping user characteristics corresponding to each historical user, the sample mapping user characteristics and multi-sample target domain content characteristics in a target domain are input into a trained content recommendation model, sample interestingness of the user to each sample target domain content in the target domain is predicted according to the sample mapping user characteristics of the historical users through each neural network of a prediction layer in the content recommendation model, click probability of each sample target domain content is predicted according to the predicted interestingness, sample click prediction scores of each sample target domain content are obtained, and the click prediction scores can also be sample click prediction values, namely click prediction results of the sample target domain content.

In one embodiment, the content recommendation model is obtained by a first stage training step, and the task-oriented mapping model is obtained by a second stage training step, the first stage training step including: acquiring historical interaction information and a corresponding click record label in a target domain; the historical interactive information comprises sample target domain user characteristics and sample target domain content characteristics corresponding to historical users in the target domain; the click record label is a click record of sample target domain content in the target domain; inputting the user characteristics of the sample target domain and the content characteristics of the sample target domain into a content recommendation model to be trained, and performing click prediction on the content of the sample target domain; determining a click prediction loss value according to a click prediction result and a click record label; and adjusting parameters of the content recommendation model according to the click prediction loss value and continuing training until the training stopping condition is met.

The server inputs the obtained sample target domain user characteristics and the sample target domain content characteristics into a content recommendation model, the click prediction result is obtained by predicting the click probability of the historical user on the sample target domain content according to the sample target domain user characteristics, and the click prediction loss value is determined according to the click prediction result and the click record label. And the server further adjusts parameters of the content recommendation model according to the click prediction loss value and continues training until the training stopping condition is met, and then the training is finished. The result of the click prediction includes a click prediction value of each sample target domain content, and specifically, a click prediction loss value may be calculated according to a difference between the click prediction value of each sample target domain content and the click record label.

In one embodiment, the server may further acquire historical interaction information and corresponding click record labels in the source domain in the same training manner, use the historical interaction content as sample source domain content, use the historical interaction information in the source domain as a training sample of a training content recommendation model, and use the click record labels as training labels to train the content recommendation model in the source domain. The training mode of the content recommendation model in the source domain is the same as the above training mode, and is not described herein again. It is understood that the content recommendation model in the target domain and the content recommendation model in the source domain may be obtained by training in the first stage.

In one embodiment, as shown in fig. 6, a specific artificial intelligence based content processing method is provided, which includes the following steps:

the content recommendation model is obtained by training in a first-stage training step, wherein the first-stage training step comprises the following steps:

s602, acquiring historical interaction information and a corresponding click record label in a target domain; the historical interactive information comprises sample target domain user characteristics and sample target domain content characteristics corresponding to historical users in the target domain; the click record label is a click record of sample target domain content in the target domain.

And S604, inputting the user characteristics of the sample target domain and the content characteristics of the sample target domain into a content recommendation model to be trained, and performing click prediction on the content of the sample target domain.

And S606, determining a click prediction loss value according to the click prediction result and the click record label.

And S608, adjusting parameters of the content recommendation model according to the click prediction loss value and continuing training until the training stopping condition is met, and obtaining the trained content recommendation model.

The task-oriented mapping model is obtained through training in a second-stage training step, wherein the second-stage training step comprises the following steps:

s610, acquiring historical behavior information of historical users shared in a source domain and a target domain and corresponding click record labels; the historical interaction information comprises sample source domain user characteristics of a historical user in a source domain and sample target domain content characteristics in a target domain; the click record label is a click record of sample target domain content in the target domain.

S612, performing feature mapping on the sample source domain user features through the task-oriented mapping model to be trained to obtain the sample mapping user features mapped from the source domain to the target domain.

And S614, inputting the sample mapping user characteristics and the sample target domain content characteristics into a pre-trained content recommendation model, and performing click prediction on the sample target domain content.

And S616, determining a task guide loss value according to the click prediction result and the click record label.

And S618, adjusting parameters of the task guide mapping model according to the task guide loss value and continuing training until the training stopping condition is met, and obtaining the trained task guide mapping model.

By taking historical users shared in a source domain and a target domain, taking sample source domain user characteristics in the source domain and sample target domain content characteristics in the target domain as training samples, taking click records corresponding to sample target domain content in the target domain as training labels, predicting corresponding task guidance based on clicking of the sample target domain content, and training by using a loss value of the task guidance, the target domain user characteristics in the target domain are avoided, errors of mapping the source domain user characteristics to the mapping user characteristics in the target domain are reduced, the accuracy of mapping the user characteristics to the target domain based on the click prediction task guidance can be effectively improved, and the accuracy of content recommendation of new users in the target domain is effectively improved.

The application also provides a specific application scenario. In particular, the parent application may be an instant messaging application, for example the parent application is a WeChat application. The WeChat application comprises a 'see-at-a-look' application, namely an article recommendation application, which is a public number content recommendation platform in the WeChat application and can recommend personalized and customized public number articles and video information streams according to the interests of users on WeChat and associated platforms. The WeChat application also comprises a video number application, namely a video recommendation application, which is a short video recommendation platform in the WeChat application and can recommend personalized and customized short video content according to the interest of a user on the WeChat and the associated platform. The look-at-a-see application may be applied as a first sub-application in the wechat application and the video number may be applied as a second sub-application in the wechat application.

For the user in the WeChat application, the user identification of the user in each sub application is consistent with the user identification in the WeChat application of the parent application. The one-look application comprises a plurality of candidate contents for recommendation, namely candidate shared contents of articles to be recommended. Similarly, the video number application includes a plurality of candidate contents for recommendation, that is, video sharing contents to be recommended. Generally, each product field has a corresponding recommendation system. The watching application comprises a pre-trained content recommendation model corresponding to the field, and the video number application also comprises a pre-trained content recommendation model corresponding to the field. The watch-at-a-glance application also includes a pre-trained task-oriented mapping model between the video number application and the watch-at-a-glance application.

For users in the WeChat application, a "see-at-a-glance" application, as well as a "video number" application, may be accessed. For example, some users do not have any interaction data, i.e., historical interaction behavior information, in an application in a certain product field. For example, for a newly added user, such as when the new user just starts using the video number application, there is no interactive behavior information in the video number application. However, the user may have interactive data in other product applications, such as a one-view application, that is, there is interactive behavior information of the user in the one-view application. At this time, the user is a new user in the one-view application, and the first sub-application of the one-view application can be used as a target domain, and the second sub-application of the video number can be used as a source domain.

When the new user accesses the application for watching at a glance, the second sub-application, namely the video number application, which runs in the parent application and has the interactive behavior information corresponding to the user identification is searched according to the user identification. And then the interactive behavior information of the user is obtained from the video number application, and the source domain user characteristics are extracted from the interactive behavior information, or the source domain user characteristics extracted in advance can be directly obtained from the video number application.

And then, carrying out feature mapping on the source domain user features in the video number application by aiming at a pre-trained task guide mapping model between the video number application and the watching-at-a-glance application to obtain mapping user features which are mapped from the video number application to the watching-at-a-glance application, so that more accurate mapping user features of a new user in the watching-at-a-glance application in a target domain can be effectively obtained. And then clicking and predicting recommended contents according to the mapping user characteristics and the target domain content characteristics of each candidate content in the one-view application by looking at the pre-trained content recommendation model in the one-view application.

Further, the content to be recommended is screened from the candidate content according to the click prediction result, and the content to be recommended screened from the watching application is pushed to the new user. Therefore, when content recommendation is performed on a new user in a view-one application target domain, the mapping user characteristics of the new user in the view-one application can be learned according to the source domain user characteristics of the new user in the video number application source domain, and content recommendation is performed by using the mapping user characteristics, so that the problem that the user characteristics of the new user in the view-one application are insufficient can be effectively solved, the content recommendation accuracy of the new user in the target domain can be effectively improved, and the cross-domain cold start recommendation effect is effectively improved.

It should be understood that although the various steps in the flow charts of fig. 2-6 are shown in order as indicated by the arrows, the steps are not necessarily performed in order as indicated by the arrows. The steps are not performed in the exact order shown and described, and may be performed in other orders, unless explicitly stated otherwise. Moreover, at least some of the steps in fig. 2-6 may include multiple steps or multiple stages, which are not necessarily performed at the same time, but may be performed at different times, which are not necessarily performed in sequence, but may be performed in turn or alternately with other steps or at least some of the other steps.

In one embodiment, as shown in fig. 7, an artificial intelligence based content recommendation apparatus 700 is provided, which may be a part of a computer device using a software module or a hardware module, or a combination of the two, and specifically includes: a data acquisition module 702, a feature mapping module 704, a content prediction module 706, and a content recommendation module 708, wherein:

a data obtaining module 702, configured to obtain source domain user characteristics corresponding to a new user of a target domain in a source domain;

the feature mapping module 704 is configured to perform feature mapping on the source domain user features through the trained task-oriented mapping model to obtain mapping user features mapped from the source domain to the target domain;

the content prediction module 706 is configured to perform click prediction on recommended content according to the mapping user characteristics and the target domain content characteristics of each candidate content in the target domain through a pre-trained content recommendation model;

a content recommending module 708, configured to filter, according to a result of the click prediction, content to be recommended from the candidate content; and pushing the content to be recommended to the new user of the target domain.

In one embodiment, the data obtaining module 702 is further configured to search, in the multiple source domains, a source domain having interaction behavior information corresponding to a user identifier according to the user identifier of the new user in the target domain; acquiring source domain user characteristics corresponding to user identification from a source domain; the source domain user characteristics are obtained by performing characteristic extraction based on the interaction behavior information of the user identification in the source domain.

In one embodiment, the feature mapping module 704 is further configured to input the source domain user features into a trained task-oriented mapping model, and perform feature matrix transformation on the source domain user features through a mapping layer of the task-oriented mapping model; and outputting the mapping user characteristics for mapping the source domain user characteristics to the target domain according to the result of the characteristic matrix transformation.

In one embodiment, the content prediction module 706 is further configured to input the mapping user characteristics and the target domain content characteristics of the candidate content in the target domain into a pre-trained content recommendation model, and predict the interest level of the new user in each candidate content through the content recommendation model; and carrying out click prediction scoring on each candidate content according to the interestingness to obtain a click prediction result corresponding to each candidate content.

In one embodiment, the data obtaining module 702 is further configured to obtain an access request corresponding to a new user in the first sub-application, where the access request includes a user identifier; the first sub-application corresponds to a target domain; the first sub-application runs in an environment provided by the parent application; searching a second sub-application running in the parent application and having interactive behavior information corresponding to the user identifier according to the user identifier; the second sub-application corresponds to the source domain; the second sub-application comprises source domain user characteristics obtained based on the interactive behavior information of the user identification.

In one embodiment, the content recommendation model is obtained by training in a first-stage training step, and the device further comprises a first model training module for obtaining historical interaction information and corresponding click record labels in the target domain; the historical interactive information comprises sample target domain user characteristics and sample target domain content characteristics corresponding to historical users in the target domain; the click record label is a click record of sample target domain content in the target domain; inputting the user characteristics of the sample target domain and the content characteristics of the sample target domain into a content recommendation model to be trained, and performing click prediction on the content of the sample target domain; determining a click prediction loss value according to a click prediction result and a click record label; and adjusting parameters of the content recommendation model according to the click prediction loss value and continuing training until the training stopping condition is met.

In one embodiment, as shown in fig. 8, another artificial intelligence based content recommendation apparatus 700 is provided, where the task-oriented mapping model is obtained through training in a second stage training step, the apparatus further includes a second model training module 7011 and a second model training module 7012, and the second model training module is configured to obtain historical interaction information and corresponding click record labels; the historical interaction information comprises sample source domain user characteristics of a historical user in a source domain and sample target domain content characteristics in a target domain; performing feature mapping on the sample source domain user features through a task-oriented mapping model to be trained to obtain sample mapping user features mapped from a source domain to a target domain; inputting the sample mapping user characteristics and the sample target domain content characteristics into a pre-trained content recommendation model, and performing click prediction on the sample target domain content; determining a task guide loss value according to a result of the click prediction and the click record label; and adjusting parameters of the task guide mapping model according to the task guide loss value and continuing training until the training stopping condition is met.

In one embodiment, as shown in fig. 9, an artificial intelligence based content processing apparatus 900 is provided, which may be a part of a computer device using software modules or hardware modules, or a combination of both, and specifically includes: a sample acquisition module 902, a feature mapping module 904, a content prediction module 906, and a parameter adjustment module 908, wherein:

a sample obtaining module 902, configured to obtain historical interaction information and a corresponding click record label; the historical interaction information comprises sample source domain user characteristics of a historical user in a source domain and sample target domain content characteristics in a target domain;

the feature mapping module 904 is configured to perform feature mapping on the sample source domain user features through a task-oriented mapping model to be trained to obtain sample mapping user features mapped from a source domain to a target domain;

the content prediction module 906 is configured to input the sample mapping user characteristics and the sample target domain content characteristics to a pre-trained content recommendation model, and perform click prediction on the sample target domain content;

a parameter adjustment module 908, configured to determine a task guidance loss value according to a result of the click prediction and the click record label; and adjusting parameters of the task guide mapping model according to the task guide loss value and continuing training until the training stopping condition is met.

In one embodiment, the sample obtaining module 902 is further configured to obtain user identifiers corresponding to historical users shared in the source domain and the target domain; respectively acquiring historical interaction information of historical users in a source domain and a target domain according to the user identification; extracting sample source domain user characteristics of a historical user in a source domain, sample target domain content characteristics of the historical user in a target domain and click record labels corresponding to the historical interaction information in the target domain from the historical interaction information; and the click record label is a click record corresponding to each sample target domain content in the historical interaction information of the target domain.

In one embodiment, the feature mapping module 904 is further configured to input the sample source-domain user features into a task-oriented mapping model to be trained, and perform feature matrix transformation on the sample source-domain user features through a mapping layer of the task-oriented mapping model; and outputting sample mapping user characteristics for mapping the sample source domain user characteristics to the target domain according to the result of the characteristic matrix transformation.

In one embodiment, the content prediction module 906 is further configured to input the sample mapping user characteristics and the sample target domain content characteristics into a pre-trained content recommendation model, and predict the sample interest degree of the historical user for each sample target domain content through the content recommendation model; and performing click prediction on the content of each sample target domain according to the interest degree of the sample to obtain a click prediction result corresponding to the content of each sample target domain.

In one embodiment, the parameter adjustment module 908 is further configured to determine a task oriented loss value according to a difference between a result of the click prediction and the click record label; and adjusting parameters of the task guide mapping model according to the task guide loss value so that the difference between the click prediction result and the click record label is reduced in the iterative training process of the task guide mapping model.

In an embodiment, as shown in fig. 10, another artificial intelligence based content processing apparatus 900 is provided, where a content recommendation model is obtained by training in a first stage training step, and a task oriented mapping model is obtained by training in a second stage training step, where the content processing apparatus 900 further includes a first model training module 901 configured to obtain historical interaction information and corresponding click record labels in a target domain; the historical interactive information comprises sample target domain user characteristics and sample target domain content characteristics corresponding to historical users in the target domain; the click record label is a click record of sample target domain content in the target domain; inputting the user characteristics of the sample target domain and the content characteristics of the sample target domain into a content recommendation model to be trained, and performing click prediction on the content of the sample target domain; determining a click prediction loss value according to a click prediction result and a click record label; and adjusting parameters of the content recommendation model according to the click prediction loss value and continuing training until the training stopping condition is met.

For specific limitations of the content processing device based on artificial intelligence, reference may be made to the above limitations of the content processing device based on artificial intelligence, which are not described herein again. The various modules in the artificial intelligence based content processing apparatus described above may be implemented in whole or in part by software, hardware, and combinations thereof. The modules can be embedded in a hardware form or independent from a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.

In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as shown in fig. 11. The computer device includes a processor, a memory, and a network interface connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of an operating system and computer programs in the non-volatile storage medium. The database of the computer equipment is used for storing data such as source domain user characteristics, candidate contents, contents to be recommended, interaction behavior information and the like. The network 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 an artificial intelligence based content processing apparatus method.

Those skilled in the art will appreciate that the architecture shown in fig. 11 is merely a block diagram of some of the structures associated with the disclosed aspects and is not intended to limit the computing devices to which the disclosed aspects apply, as particular computing devices may include more or less components than those shown, or may combine certain components, or have a different arrangement of components.

In one embodiment, a computer device is further provided, which includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method embodiments when executing the computer program.

In an embodiment, a computer-readable storage medium is provided, in which a computer program is stored which, when being executed by a processor, carries out the steps of the above-mentioned method embodiments.

In one embodiment, a computer program product or computer program is provided that includes computer instructions stored in a computer-readable storage medium. The computer instructions are read by a processor of a computer device from a computer-readable storage medium, and the computer instructions are executed by the processor to cause the computer device to perform the steps in the above-mentioned method embodiments.

It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by hardware instructions of a computer program, which can be stored in a non-volatile computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. Any reference to memory, storage, database or other medium used in the embodiments provided herein can include at least one of non-volatile and volatile memory. Non-volatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical storage, or the like. Volatile Memory can include Random Access Memory (RAM) or external cache Memory. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM), among others.

The technical features of the above embodiments can be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the above embodiments are not described, but should be considered as the scope of the present specification as long as there is no contradiction between the combinations of the technical features.

The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims

1. An artificial intelligence based content recommendation method, the method comprising:

performing feature matrix transformation on the source domain user features through a trained task-oriented mapping model, and obtaining mapping user features mapped from the source domain to the target domain according to the transformed vector representation in the feature space of the target domain; the task-oriented mapping model is obtained by performing click prediction on the content characteristics of the sample target domain based on the user characteristics of the sample source domain, and is a machine learning model with the capability of mapping the user characteristics of the source domain to the characteristic space of the target domain;

through a trained content recommendation model, click prediction of recommended content is carried out according to the mapping user characteristics and the target domain content characteristics of each candidate content in the target domain; the content recommendation model is obtained by training based on the user characteristics of the sample target domain and the content characteristics of the sample target domain, and is a machine learning model with the capability of performing click prediction on candidate content according to the user characteristics in the target domain;

and pushing the content to be recommended to the new user of the target domain.

2. The method of claim 1, wherein the obtaining of the source domain user characteristics corresponding to the new user of the target domain in the source domain comprises:

according to the user identification of the new user in the target domain, searching a source domain with interactive behavior information corresponding to the user identification in a plurality of source domains;

acquiring source domain user characteristics corresponding to the user identification from the source domain; the source domain user characteristics are obtained by performing characteristic extraction based on the interactive behavior information of the user identification in the source domain.

3. The method of claim 1, further comprising:

acquiring an access request corresponding to a new user in a first sub-application, wherein the access request comprises a user identifier; the first sub-application corresponds to the target domain; the first sub-application runs in an environment provided by a parent application;

searching a second sub-application running in the parent application and having interactive behavior information corresponding to the user identifier according to the user identifier; the second sub-application corresponds to the source domain; and the second sub-application comprises source domain user characteristics obtained based on the interactive behavior information of the user identification.

4. The method according to any one of claims 1 to 3, wherein the content recommendation model is obtained by a first stage training step, the first stage training step comprising:

acquiring historical interaction information and a corresponding click record label in the target domain; the historical interaction information comprises sample target domain user characteristics and sample target domain content characteristics corresponding to historical users in the target domain; the click record label is a click record of sample target domain content in the target domain;

inputting the sample target domain user characteristics and the sample target domain content characteristics into a content recommendation model to be trained, and performing click prediction on sample target domain content;

determining a click prediction loss value according to a click prediction result and the click record label;

and adjusting parameters of the content recommendation model according to the click prediction loss value and continuing training until the training stopping condition is met.

5. The method according to any one of claims 1 to 3, wherein the task-oriented mapping model is obtained by a second-stage training step, the second-stage training step comprising:

6. A method for artificial intelligence based content processing, the method comprising:

performing feature matrix transformation on the sample source domain user features through a task-oriented mapping model to be trained, and obtaining sample mapping user features mapped from the source domain to the target domain according to the transformed vector representation in the feature space of the target domain;

inputting the sample mapping user characteristics and the sample target domain content characteristics into a trained content recommendation model, and performing click prediction on sample target domain content; the content recommendation model is obtained by training based on the user characteristics of the sample target domain and the content characteristics of the sample target domain, and is a machine learning model with the capability of performing click prediction on candidate content according to the user characteristics in the target domain;

adjusting parameters of the task guide mapping model according to the task guide loss value and continuing training until the training stopping condition is met, and obtaining a trained task guide mapping model; the trained task-oriented mapping model is a machine learning model with the capability of predicting candidate contents by clicking according to the user characteristics in the target domain, and is used for performing characteristic mapping on the source domain user characteristics of the new user in the target domain in order to recommend the contents in the target domain to the new user according to the obtained mapping user characteristics.

7. The method of claim 6, wherein obtaining historical interaction information and corresponding click record tags comprises:

acquiring user identifications corresponding to history users shared in the source domain and the target domain;

respectively acquiring historical interaction information of the historical user in the source domain and the target domain according to the user identification;

extracting sample source domain user characteristics of the historical users in the source domain, sample target domain content characteristics of the historical users in the target domain and click record labels corresponding to the historical interaction information in the target domain from the historical interaction information; and the click record label is a click record corresponding to each sample target domain content in the historical interaction information of the target domain.

8. The method of claim 6, wherein said characterizing the sample source-domain user features by a task-oriented mapping model to be trained to obtain sample mapped user features that map from the source domain to the target domain, comprises:

inputting the sample source domain user characteristics into a task-oriented mapping model to be trained, and performing characteristic matrix transformation on the sample source domain user characteristics through a mapping layer of the task-oriented mapping model;

and outputting sample mapping user characteristics for mapping the sample source domain user characteristics to the target domain according to the result of the characteristic matrix transformation.

9. The method of claim 6, wherein inputting the sample mapped user features and the sample target domain content features into a pre-trained content recommendation model for click prediction of sample target domain content comprises:

inputting the sample mapping user characteristics and the sample target domain content characteristics into a pre-trained content recommendation model, and predicting the sample interest degree of the historical user on the content of each sample target domain through the content recommendation model;

and performing click prediction on the content of each sample target domain according to the sample interest degree to obtain a click prediction result corresponding to the content of each sample target domain.

10. The method of claim 6, wherein determining a task oriented loss value based on the result of the click prediction and the click record label comprises:

determining a task guide loss value according to the difference between the click prediction result and the click record label;

adjusting parameters of the task-oriented mapping model according to the task-oriented loss value and continuing training, comprising:

and adjusting parameters of the task guide mapping model according to the task guide loss value so as to reduce the difference between the click prediction result and the click record label in the iterative training process of the task guide mapping model.

11. The method according to any one of claims 6 to 10, wherein the content recommendation model is obtained by a first stage training step, and the task-oriented mapping model is obtained by a second stage training step, and the first stage training step comprises:

12. An artificial intelligence based content recommendation apparatus, characterized in that the apparatus comprises:

the feature mapping module is used for performing feature matrix transformation on the source domain user features through a trained task-oriented mapping model, and obtaining mapping user features mapped from the source domain to the target domain according to the transformed vector representation in the feature space of the target domain; the task-oriented mapping model is obtained by performing click prediction on the content characteristics of the sample target domain based on the user characteristics of the sample source domain, and is a machine learning model with the capability of mapping the user characteristics of the source domain to the characteristic space of the target domain;

the content prediction module is used for performing click prediction on recommended content according to the mapping user characteristics and the target domain content characteristics of each candidate content in the target domain through a trained content recommendation model; the content recommendation model is obtained by pre-training based on the sample target domain user characteristics and the sample target domain content characteristics, and is a machine learning model with the capability of clicking and predicting candidate content according to the user characteristics in the target domain;

13. An artificial intelligence based content processing apparatus, the apparatus comprising:

the feature mapping module is used for performing feature matrix transformation on the sample source domain user features through a task-oriented mapping model to be trained, and obtaining sample mapping user features mapped from the source domain to the target domain according to the transformed vector representation in the feature space of the target domain;

the content prediction module is used for inputting the sample mapping user characteristics and the sample target domain content characteristics into a trained content recommendation model to carry out click prediction on the sample target domain content; the content recommendation model is obtained by training based on the user characteristics of the sample target domain and the content characteristics of the sample target domain, and is a machine learning model with the capability of performing click prediction on candidate content according to the user characteristics in the target domain;

the parameter adjusting module is used for determining a task guide loss value according to a click prediction result and the click record label; adjusting parameters of the task guide mapping model according to the task guide loss value and continuing training until the training stopping condition is met, and obtaining a trained task guide mapping model; the trained task-oriented mapping model is a machine learning model with the capability of predicting candidate contents by clicking according to the user characteristics in the target domain, and is used for performing characteristic mapping on the source domain user characteristics corresponding to the new user in the target domain so as to recommend the contents in the target domain to the new user according to the obtained mapping user characteristics.

14. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor realizes the steps of the method of any one of claims 1 to 11 when executing the computer program.

15. A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the steps of the method according to any one of claims 1 to 11.