CN112102015B

CN112102015B - Article recommendation method, meta-network processing method, device, storage medium and equipment

Info

Publication number: CN112102015B
Application number: CN202011282911.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: 2020-11-17
Filing date: 2020-11-17
Publication date: 2021-03-02
Anticipated expiration: 2040-11-17
Also published as: CN112102015A

Abstract

The application relates to an article recommendation method, a meta-network processing device, a storage medium and equipment. The item recommendation method comprises the following steps: acquiring attribute information of an article and an object identifier of an article interaction object; extracting the characteristics of the attribute information through a meta-stretch network to obtain a stretch vector; extracting the characteristics of the object identification through a meta-offset network to obtain an offset vector; performing feature crossing on the vectorized article identification of the article and the stretching vector, and fusing the obtained crossing feature vector and the offset vector to obtain a fused feature vector; determining a recommendation probability based on the fused feature vector, the vectorized attribute information, and vectorized object information of a target object; and recommending the item to the target object when the recommendation probability meets a recommendation condition. By the method, the article recommendation accuracy can be improved, and therefore the accurate recommendation effect is achieved.

Description

Article recommendation method, meta-network processing method, device, storage medium and equipment

Technical Field

The present application relates to the field of artificial intelligence technologies, and in particular, to an article recommendation method, a meta-network processing method, an apparatus, a storage medium, and a device.

Background

With the development of internet technology, users can acquire various goods, such as application programs or other service products, from the internet. For an item provider, a new item to be developed needs to be accurately recommended to a user with a demand, so as to improve the product utilization rate.

Since the developed new items are generally used by only a small number of users, when recommendation is performed, the traditional recommendation scheme not only recommends interested and mature items to the users, but also pushes the new items to the users, so as to improve the utilization rate of the new items. However, in the above recommendation scheme, the user is interested in the mature item, but not necessarily in the new item, thereby resulting in low accuracy of recommending the item to the user.

Disclosure of Invention

In view of the foregoing, it is desirable to provide an item recommendation method, a meta-network processing method, an apparatus, a storage medium, and a device, which can improve the accuracy of item recommendation.

A method of item recommendation, the method comprising:

acquiring attribute information of an article and an object identifier of an article interaction object;

extracting the characteristics of the attribute information through a meta-stretch network to obtain a stretch vector; extracting the characteristics of the object identification through a meta-offset network to obtain an offset vector;

performing feature crossing on the vectorized article identification of the article and the stretching vector, and fusing the obtained crossing feature vector and the offset vector to obtain a fused feature vector;

determining a recommendation probability based on the fused feature vector, the vectorized attribute information, and vectorized object information of a target object;

and recommending the item to the target object when the recommendation probability meets a recommendation condition.

In one embodiment thereof, the method further comprises:

inputting the training fusion feature vector, the vectorized training attribute information and the vectorized training object information of the object sample into a recommendation model;

processing the training fusion feature vector, the vectorized training attribute information and the vectorized training object information of the object sample through the recommendation model to obtain a training recommendation probability;

and when the training recommendation probability meets a training condition, executing the step of updating the network parameters in the meta stretch network and the meta offset network respectively according to the first loss value.

In one embodiment thereof, the method further comprises:

vectorizing the training article identification through an embedding layer to obtain a training article identification embedding vector;

calculating a second loss value between the training article identification embedding vector and the sample label;

and when the training recommendation probability meets the training condition, updating the network parameters in the embedded layer according to the second loss value.

An item recommendation device, the device comprising:

the acquisition module is used for acquiring the attribute information of the article and the object identifier of the article interaction object;

the characteristic extraction module is used for extracting the characteristics of the attribute information through a meta-stretch network to obtain a stretch vector; extracting the characteristics of the object identification through a meta-offset network to obtain an offset vector;

the characteristic processing module is used for carrying out characteristic crossing on the vectorized article identification of the article and the stretching vector and fusing the obtained crossed characteristic vector with the offset vector to obtain a fused characteristic vector;

a determination module for determining a recommendation probability based on the fused feature vector, the vectorized attribute information, and the vectorized object information of the target object;

and the recommending module is used for recommending the object to the target object when the recommending probability meets a recommending condition.

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:

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

According to the article recommendation method, the article recommendation device, the computer equipment and the storage medium, the attribute information of the article and the object identification of the article interaction object are respectively subjected to feature extraction through the meta stretch network and the meta offset network to obtain the corresponding stretch vector and the offset vector, then the article identification of the article is processed according to the stretch vector and the offset vector to obtain the fusion feature vector, and the recommendation probability is determined based on the fusion feature vector, the vectorized attribute information and the vectorized object information of the target object, so that even if the article of the cold door is used by a small number of users, accurate recommendation of the cold door article can be realized through the article identification, the attribute information and the object identification of the small number of article interaction objects. In addition, after the vectorized article identification of the article is subjected to feature intersection with the stretching vector, the obtained intersection feature vector is fused with the offset vector, so that the noise influence on the feature of the article identification is reduced, and further, the article recommendation is performed by utilizing the fusion feature vector of the article identification, so that the article recommendation accuracy can be further improved.

A meta-network processing method, the method comprising:

acquiring training attribute information of an article sample and a training object identifier of an object sample;

extracting the characteristics of the training attribute information through a meta-stretch network to be trained to obtain a training stretch vector; performing feature extraction on the training object identification through a to-be-trained meta-offset network to obtain a training offset vector;

performing feature crossing on the vectorized training article identification of the article sample and the training stretching vector, and fusing the obtained training cross feature vector and the training offset vector to obtain a training fusion feature vector;

calculating a first loss value between the training fused feature vector and a sample label;

and updating the network parameters in the meta stretch network and the meta offset network respectively according to the first loss value.

A meta-network processing apparatus, the apparatus comprising:

the acquisition module is used for acquiring training attribute information of the article sample and a training object identifier of the object sample;

the feature extraction module is used for extracting features of the training attribute information through a meta-stretch network to be trained to obtain a training stretch vector; performing feature extraction on the training object identification through a to-be-trained meta-offset network to obtain a training offset vector;

the characteristic processing module is used for performing characteristic crossing on the vectorized training article identification of the article sample and the training stretching vector, and fusing the obtained training cross characteristic vector and the training offset vector to obtain a training fusion characteristic vector;

the calculation module is used for calculating a first loss value between the training fusion feature vector and a sample label;

and the updating module is used for respectively updating the network parameters in the meta stretch network and the meta offset network according to the first loss value.

According to the meta-network processing method, the meta-network processing device, the computer equipment and the storage medium, feature extraction is performed on training attribute information of an article sample through a meta-stretch network to be trained, and a training stretch vector is obtained; and extracting the feature of the training object mark of the object sample through a meta-bias network to be trained to obtain a training bias vector, performing feature processing on the training object mark of the object sample which is vectorized based on the training stretching vector and the training bias vector to obtain a training fusion feature vector related to the training object mark, continuously updating network parameters in the meta-stretching network and the meta-bias network through a first loss value between the training fusion feature vector and a sample label even if the object sample is a cold-door object, so that the meta-stretching network and the meta-bias network continuously learn the feature vector related to the training object mark and fused with the object attribute and the object mark of the interactive object, and obtaining the mark feature vector of the cold-door object through the trained meta-stretching network and the meta-bias network even if the cold-door object is recommended and the cold-door object is used by only a small number of users, based on the identification feature vector, the attribute information and the object identification of a small number of object interaction objects, accurate recommendation of cold objects can be achieved.

Drawings

FIG. 1 is a diagram of an exemplary environment in which a method for recommending items may be implemented;

FIG. 2 is a flow diagram that illustrates a method for item recommendation, according to one embodiment;

FIG. 3 is a schematic diagram of a model structure in one embodiment;

FIG. 4 is a flowchart illustrating the steps of training a meta stretch network and a meta shift network in one embodiment;

FIG. 5 is a flowchart illustrating the step of updating network parameters in the embedding layer according to another embodiment;

FIG. 6 is a flow diagram illustrating a meta-network processing method in accordance with another embodiment;

FIG. 7 is a schematic view of a model structure in another embodiment;

FIG. 8 is a block diagram showing the structure of an article recommendation apparatus according to an embodiment;

FIG. 9 is a block diagram showing the construction of an article recommending apparatus according to another embodiment;

FIG. 10 is a block diagram of a meta-network processing device in one embodiment;

FIG. 11 is a block diagram showing the structure of a meta-network processing apparatus according to another embodiment;

FIG. 12 is a diagram illustrating an internal structure of a computer device according to an 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.

Before describing the present application in detail, the techniques employed in the present application will first be briefly described:

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.

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

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. The port opening and closing state characteristics corresponding to the target equipment can be classified through a machine learning method, and therefore the type of the target equipment is determined.

Cloud technology (Cloud technology) refers to a hosting technology for unifying serial resources such as hardware, software, and network in a wide area network or a local area network to realize calculation, storage, processing, and sharing of data.

Cloud technology (Cloud technology) is based on a general term of network technology, information technology, integration technology, management platform technology, application technology and the like applied in a Cloud computing business model, can form a resource pool, is used as required, and is flexible and convenient. Cloud computing technology will become an important support. Background services of the technical network system require a large amount of computing and storage resources, such as video websites, picture-like websites and more web portals. With the high development and application of the internet industry, each article may have its own identification mark and needs to be transmitted to a background system for logic processing, data in different levels are processed separately, and various industrial data need strong system background support and can only be realized through cloud computing.

By combining artificial intelligence technology with cloud technology, artificial intelligence cloud services, also commonly referred to as AI as a Service (AI as a Service), can be provided. The method is a service mode of an artificial intelligence platform, and particularly, the AIaaS platform splits several types of common AI services and provides independent or packaged services at a cloud. This service model is similar to the one opened in an AI theme mall: all developers can access one or more artificial intelligence services provided by the platform through an API (application programming interface), and part of the qualified developers can also use an AI framework and an AI infrastructure provided by the platform to deploy and operate and maintain the dedicated cloud artificial intelligence services. In the application, the trained meta-stretch network and the trained meta-offset network are deployed in a recommendation system, and when article recommendation is needed, the meta-stretch network and the meta-offset network in the recommendation system can be called to preheat (Warm up) the article identifier of the cold article, so that a fusion feature vector of the article identifier of the cold article is obtained, and the fusion feature vector is the feature vector (Warm ID embedding) of the preheated article identifier.

The scheme provided by the embodiment of the application relates to technologies such as cloud technology and machine learning of artificial intelligence, and is specifically explained by the following embodiments:

the item recommendation method provided by the application can be applied to the application environment shown in fig. 1. In the application environment, the terminal 102 and the server 104 are included, and the terminal 102 and the server 104 may be connected through communication connection manners such as bluetooth, USB (Universal Serial Bus), or a network, which is not limited herein.

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. Various applications for recommending items can be installed in the terminal 102, for example, the applications can be a social application, a shopping application, a video application, or the like, and a recommendation system is built in the applications, and the corresponding items can be recommended to the user through the recommendation system.

The server 104 may be an independent physical server, may also be a server cluster composed of a plurality of physical servers, and may be a cloud server that provides basic cloud computing services such as a cloud service, a cloud database, cloud computing, a cloud function, cloud storage, a Network service, cloud communication, a middleware service, a domain name service, a security service, a Content Delivery Network (CDN), a big data and artificial intelligence platform, and the like. The server 104 may be installed with a recommendation system, and the terminal sends the item identifier to be recommended to the server 104, and the server may invoke the recommendation system, so that the recommendation system recommends the corresponding item to the user based on the item identifier.

In one embodiment, as shown in fig. 2, an item recommendation method is provided, which may be executed by the terminal 102 or the server 104 in fig. 1, or may be executed by the terminal 102 and the server 104 cooperatively. The method is described as an example executed by the server 104, and comprises the following steps:

s202, acquiring attribute information of the article and object identification of the article interaction object.

The item may refer to an item that needs to be recommended to the target object, that is, when the calculated recommendation probability is greater than the probability threshold, the item is recommended to the target object. In practical applications, the item may refer to a cold item that needs to be recommended to the target object. The cold item may be an item that is used less by a user, such as a commodity that is used less by a user; alternatively, the cold item may be a new item that has just been marketed, such as an advertisement that has just been pushed.

The attribute information may be information describing attributes and functions of the article, including at least one of: the source of the item, the manufacturer, the item website, the use, the performance, the composition, the user-oriented group, and other item information, among others. For example, for the advertisement just pushed out, the attribute information may be information of the user group to which the advertisement is directed and the item to be promoted, and the like.

An interactive item object may refer to an object that has an interaction with the item, such as a user that has used a certain item, or a user that has clicked on a certain item link, or a user that has viewed a certain advertisement. The object identification may refer to an identification for uniquely identifying the interactive item object, such as identity information of the user.

In one embodiment, when receiving an item recommendation instruction, a server obtains an item identifier and attribute information of a corresponding item from an item library according to the item recommendation instruction, and then obtains an object identifier of an interactive item object according to the item identifier.

In another embodiment, the server extracts the item identification and attribute information of the item and the object identification of the item interaction object from the item recommendation instruction when receiving the item recommendation instruction.

And S204a, extracting the characteristics of the attribute information through the meta-stretch network to obtain a stretch vector.

The meta-stretch network may refer to a meta-network obtained based on meta-learning and used for feature extraction of attribute information of an item, and through the meta-stretch network, an item identifier may be mapped from a data set to a stretch vector of a specific model space. Specifically, the meta-network is obtained by training an initial meta-stretch network based on a meta-learning method, and the training speed of the initial meta-stretch network can be increased by the meta-learning method, so that the meta-stretch network can be converged quickly, and a network for extracting the characteristics of the attribute information of the article can be obtained quickly. For example, an initial meta-stretch network is trained based on a method for predicting a gradient, and gradient prediction is performed during the training process, so that the training process can be accelerated to obtain a trained meta-stretch network.

Functionally, the stretch vector may be used as a pull-up function to map an item identification for the item from the current data set to a particular model space. For example, for a cold door item, the vectorized item identification may be converted to a rom ID embedding by the stretch vector.

In one embodiment, the server performs vectorization processing on the attribute information, and then performs feature extraction on the obtained attribute embedded vector to obtain a stretching vector.

Specifically, the server carries out vectorization processing on the attribute information through the embedding layer to obtain an attribute embedding vector; and then inputting the attribute embedded vector into a meta-stretch network, and processing the attribute embedded vector through the meta-stretch network to obtain a stretch vector. The embedded layer may refer to a layer for embedding attribute information, an item identifier, an object identifier of an item interaction object, and object information of a target object. The embedding means that attribute information, article identification, object identification of an article interaction object and object information of a target object are represented by a dense vector, so that related information is more fully learned.

In One embodiment, the server may first encode the attribute information by using One Hot encoding to obtain an attribute encoding vector; and then inputting the attribute coding vector into an embedding layer, and performing densification representation on the attribute coding vector through the embedding layer to obtain an attribute embedding vector.

For example, the server vectorizes each attribute information of the cold goods by the embedding layer to obtain an attribute embedding vector (itemfeaturembeddings) of the attribute information of the cold goods, and the attribute embedding vector is recorded as a vector

Embedding attributes into vectors

The input element stretching network can output a stretching vector

Wherein, in the step (A),

，

an initial embedded vector of item identifications representing the ith cold-door item, and

an attribute embedding vector representing the corresponding ith cold door item,

the stretching parameters are shown.

S204b, extracting the characteristics of the object identification through the meta-offset network to obtain an offset vector.

The meta-offset network can be a meta-network for extracting features of object identifiers of item interaction objects based on meta-learning, and an offset vector for reinforcing the arm ID embedding can be obtained through the meta-offset network. The meta-network can be obtained by training an initial meta-offset network based on a meta-learning method, the training speed of the initial meta-offset network can be increased by the meta-learning method, the meta-offset network can be rapidly converged, and therefore the network for extracting the characteristics of the object identification of the object interaction object can be rapidly obtained.

Functionally, the stretch vector may be used as a function of an offset to enhance the identification of items mapped from the current data set to a particular model space to remove noise effects. For example, for a cold door article, after converting vectorized article identification into a war ID embedding by a stretching vector, the war ID embedding is strengthened, so that the war ID embedding is more stable, and the influence is reduced.

In one embodiment, the server performs vectorization processing on the object identifier, and then performs feature extraction on the obtained object identifier embedded vector to obtain an offset vector.

Specifically, the server carries out vectorization processing on the object identifier through the embedded layer to obtain an object identifier embedded vector; and then inputting the object identification embedded vector into a meta-offset network, and processing the object identification embedded vector through the meta-offset network to obtain an offset vector.

For example, the server vectorizes the object identifier of each article interaction object interacted with the cold article through the embedding layer to obtain the object identifier embedding of the article interaction objectVectors (UserIDembeddings), the object id embedded vector denoted as a vector

Embedding object identification into a vector

The input element offset network can output an offset vector

Wherein, in the step (A),

，

representing an offset parameter.

S206, performing feature crossing on the vectorized article identification and the stretching vector of the article, and fusing the obtained crossed feature vector and the offset vector to obtain a fused feature vector.

Wherein the item identification may refer to an identification for uniquely identifying the item. For example, when the item is a hardware device, the item identification may be a hardware serial number or a physical address of the item; as another example, when the item is an advertisement, the item identification may be an advertisement index for the unique advertisement.

In one embodiment, when the server obtains the item identifier of the item, the server performs vectorization processing on the item identifier to obtain a vectorized item identifier, and then executes S206.

Specifically, the server inputs the article identifier into the embedding layer, and vectorizes the article identifier through the embedding layer to obtain an article identifier embedding vector. The step of performing feature intersection on the vectorized article identifier and the stretching vector of the article may specifically include: and the server performs feature crossing on the article identification embedding vector and the stretching vector.

In One embodiment, the server may first encode the item identifier by using One Hot encoding to obtain an item encoding vector; and then inputting the article coding vector into an embedding layer, and performing densification representation on the article coding vector through the embedding layer to obtain an article identification embedding vector.

As an example, the server vectorizes each attribute information of the cold goods through the embedding layer to obtain an attribute embedding vector (featurembeddings) of the attribute information of the cold goods, and the attribute embedding vector is recorded as a vector

Embedding attributes into vectors

The input element stretching network can output a stretching vector

Wherein, in the step (A),

(ii) a The vectorized article identifier of the cold-door article is subjected to dot multiplication by taking the stretching vector as a stretching function, so that an initial war ID embedding (namely, the cross feature vector) of the cold-door article can be obtained, and the expression of the initial war ID embedding is as follows:

then, the server vectorizes the object identifier of each article interaction object interacted with the cold article through the embedding layer to obtain an object identifier embedding vector (UserIDEmbeddings) of the article interaction object, and the object identifier embedding vector is marked as a vector

Embedding object identification into a vector

Input deviceThe meta-offset network can output an offset vector

Wherein, in the step (A),

，

representing an offset parameter; fusing the initial war ID embedding of the cold door article by taking the offset vector as an offset function to obtain final war ID embedding, wherein the expression of the war ID embedding is as follows:

s208, determining recommendation probability based on the fusion feature vector, the vectorized attribute information and the vectorized object information of the target object.

Wherein the target object may refer to a user who may be interested in the item, or a potential user of the item. The object information may be information for describing the target object, including target object identification and target object characteristic information. The target object identification can be identity information or a user name of the target object; the target object feature information includes but is not limited to: the age, sex, place of departure, place of work, school calendar, income, home background, etc. of the target subject. The recommendation probability is used to represent the probability of recommending the item to the target user.

In one embodiment, a server obtains object information of a target object; and vectorizing the object information through the embedding layer to obtain an object embedding vector. S208 may specifically include: the server determines a recommendation probability based on the fused feature vector, the attribute embedding vector, and the object embedding vector.

In One embodiment, the server may first encode the object information of the target object by using One Hot encoding to obtain an object encoding vector; and then inputting the object coding vector into an embedding layer, and performing densification representation on the object coding vector through the embedding layer to obtain an object embedding vector.

The server inputs the fused feature vector, the attribute embedding vector, and the object embedding vector into a recommendation model, whereby the recommendation model determines a recommendation probability based on the fused feature vector, the attribute embedding vector, and the object embedding vector.

In an embodiment, when the object information includes the target object identifier and the target object feature information, the step of performing vectorization processing on the object information through the embedding layer to obtain the object embedding vector may specifically include: the server carries out vectorization processing on the target object identification through the embedding layer to obtain a target object identification embedding vector; vectorizing the characteristic information of the target object through the embedding layer to obtain a characteristic embedding vector of the target object; wherein the object embedding vector comprises a target object identification embedding vector and a target object feature embedding vector.

Specifically, the server inputs the fusion feature vector, the attribute embedded vector, the target object identification embedded vector and the target object characteristic embedded vector into the recommendation model, so that the recommendation model calculates the fusion feature vector, the attribute embedded vector, the target object identification embedded vector and the target object characteristic embedded vector to obtain recommendation probability.

And S210, recommending the object to the target object when the recommendation probability meets the recommendation condition.

The recommendation condition may represent whether the recommendation probability reaches a condition for recommending an item to the target object, for example, when the recommendation probability is greater than or equal to a probability threshold, recommending an item to the target object; and when the recommendation probability is smaller than the probability threshold value, not recommending the object to the target object.

In order to more intuitively understand the solution of the present application, the above item recommendation method is described with reference to fig. 3: firstly, information related to an article and object information of a target object are obtained, wherein the information related to the article comprises article identification and attribute information of the article, and object identification of an article interaction object interacted with the article. And respectively carrying out vectorization processing on the object identification, the article identification, the attribute information and the object information of the target object through the embedding layer to sequentially obtain a corresponding object identification embedding vector, an article identification embedding vector, an attribute embedding vector and an object embedding vector. Then embedding the attribute into a vector input element stretching network, and outputting a stretching vector; and embedding the object identification into a vector input element offset network and outputting an offset vector. Processing the article identification embedded vector by taking the stretching vector as a stretching function to obtain a cross characteristic vector related to the article identification; and processing the crossed feature vector by taking the offset vector as an offset function to obtain a fusion feature vector related to the article identifier. And finally, inputting the fusion characteristic vector, the attribute embedded vector and the object embedded vector into a recommendation model, predicting recommendation probability through the recommendation model, and determining whether to recommend the object to the target object according to the recommendation probability.

In the above embodiment, the attribute information of the article and the object identifier of the article interaction object are respectively subjected to feature extraction through the meta stretch network and the meta shift network to obtain corresponding stretch vectors and shift vectors, then the article identifier of the article is processed according to the stretch vectors and the shift vectors to obtain fusion feature vectors, and the recommendation probability is determined based on the fusion feature vectors, the vectorized attribute information, and the vectorized object information of the target object, so that even if the article at the cold door is used by only a small number of users, accurate recommendation of the article at the cold door can be realized through the article identifier of the article, the attribute information, and the object identifier of the small number of article interaction objects. In addition, after the vectorized article identification of the article is subjected to feature intersection with the stretching vector, the obtained intersection feature vector is fused with the offset vector, so that the noise influence on the feature of the article identification is reduced, and further, the article recommendation is performed by utilizing the fusion feature vector of the article identification, so that the article recommendation accuracy can be further improved.

As shown in fig. 4, in an embodiment, the meta stretch network and the meta bias network are obtained by performing meta network processing on the meta stretch network and the meta bias network to be trained, respectively; wherein, the meta-network processing is also to train the meta-network.

Therefore, before recommending the article based on the meta-network, training the stretching metafunction, the offset metafunction and the embedding layer to be trained, wherein the training step comprises the following steps:

s402, obtaining training attribute information of the article sample and a training object identification of the object sample.

The item sample can refer to an item used for meta-network training, and the item sample has interacted with a larger number of item interaction object samples. Therefore, the recommendation probability of the item sample recommending to the item interaction object sample is a known quantity, and the known recommendation probability can be used as a label.

The training attribute information may refer to attribute information used in training phase for meta-network training, and may be information describing attributes and functions of the article. The training attribute information may specifically include at least one of: the source of the item, the manufacturer, the item website, the use, the performance, the composition, the user-oriented group, and other item information, among others.

The object sample includes: an item interaction object sample with which the item sample has interacted, and a non-item interaction object sample with which the item sample has not interacted. For the sample of article interaction objects, for example, a user using a certain article, or a user clicking a certain article link, or a user watching a certain advertisement. The object identification may refer to an identification for uniquely identifying the interactive item object, such as identity information of the user. In the training process, the article sample and the corresponding object sample may be divided into a plurality of sub-article samples and corresponding sub-object samples, and then training is performed using training data of each sub-article sample and sub-object sample.

In the initial stage of article recommendation, the article sample has a cold-start recommendation (cold-start recommendation) process, and training is performed by using the training attribute information, the training article identifier and the training object identifier of the object sample in the cold-start recommendation process of the article sample. Cold start recommendations refer to how to accurately recommend cold items to interested users, where a cold item may refer to a new item at an early stage of the market. It is noted that in the cold start recommendation process, the item samples belong to cold items. As the number of item interaction objects interacting with the item sample increases, the item sample will not belong to a cold item.

In one embodiment, when receiving an item recommendation instruction, a server obtains training item identifiers and training attribute information of corresponding item samples from an item library according to the item recommendation instruction, and then obtains training object identifiers of interactive item object samples according to the training item identifiers.

S404a, extracting the feature of the training attribute information through the meta-stretching network to be trained to obtain the training stretching vector.

The meta-stretch network to be trained refers to an untrained network model, and the meta-stretch network can perform feature extraction on training attribute information of the article sample in the training process. Functionally, the training stretch vector may be used as a pull-up function to map the training item identification of the item sample from the current data set to a particular model space. For example, for a cold door item, the vectorized training item identification may be converted to a rom ID embedding by the training stretch vector.

In one embodiment, the server performs vectorization processing on the training attribute information, and then performs feature extraction on the obtained training attribute embedded vector to obtain a training stretching vector.

Specifically, the server carries out vectorization processing on the training attribute information through an embedding layer to obtain a training attribute embedding vector; and then inputting the training attribute embedded vector into a meta-stretch network to be trained, and processing the training attribute embedded vector through the meta-stretch network to obtain a training stretch vector. The embedded layer may be used to embed the training attribute information, the training article identifier, the training object identifier, and the training object information. The embedding processing refers to representing the training attribute information, the training article identifier, the training object identifier and the training object information by dense vectors, so that the related information is more fully learned.

In One embodiment, the server may first encode the training attribute information in a One Hot encoding manner to obtain a training attribute encoding vector; and then inputting the training attribute coding vector into an embedding layer, and performing densification representation on the training attribute coding vector through the embedding layer to obtain the training attribute embedding vector.

For example, the server vectorizes each training attribute information of the article sample through the embedding layer to obtain a training attribute embedding vector of the training attribute information, and the training attribute embedding vector is recorded as a vector

Embedding training attributes into vectors

Inputting the meta stretch network to be trained, namely outputting a stretch vector

Wherein, in the step (A),

，

the initial embedded vector of training article identifications representing the ith article sample, and

a training attribute embedding vector corresponding to the ith item sample is represented,

the stretching parameters are shown.

S404b, extracting the feature of the training object mark through the meta-offset network to be trained, and obtaining the training offset vector.

The meta-bias network to be trained may refer to an untrained network model, and the meta-bias network may perform feature extraction on a training object identifier of an object sample in a training process. Functionally, the training offset vector may be used as an offset function to enhance the training item identification mapping from the current data set to a particular model space to remove noise effects.

In one embodiment, the server performs vectorization processing on the training object identifier, and then performs feature extraction on the obtained training object identifier embedded vector to obtain a training offset vector.

Specifically, the server carries out vectorization processing on the training object identification through the embedding layer to obtain a training object identification embedding vector; and then inputting the training object identification embedded vector into a meta-offset network, and processing the training object identification embedded vector through the meta-offset network to obtain a training offset vector.

For example, the server vectorizes the training object id of each object sample interacted with the article sample through the embedding layer to obtain a training object id embedding vector of the object sample, and the training object id embedding vector is recorded as a vector

Embedding training object identification into vector

The input element offset network can output a training offset vector

Wherein, in the step (A),

。

s406, performing feature crossing on the vectorized training article identification and the training stretching vector of the article sample, and fusing the obtained training cross feature vector and the training offset vector to obtain a training fusion feature vector.

Wherein, the training article identifier may refer to an identifier for uniquely identifying the article. For example, when the item sample is a hardware device, the training item identifier may be a hardware serial number or a physical address of the item sample; for another example, when the sample of items is an advertisement, the training item identification may be an advertisement index of the unique advertisement.

In one embodiment, when the server obtains the training article identifier of the article sample, the server performs vectorization processing on the training article identifier to obtain a vectorized training article identifier, and then executes S406.

Specifically, the server inputs the training article identifier into the embedding layer, and vectorizes the training article identifier through the embedding layer to obtain a training article identifier embedding vector. The step of performing feature crossing on the vectorized training article identifier and the training stretching vector of the article sample may specifically include: and the server performs feature crossing on the training article identification embedding vector and the training stretching vector.

In One embodiment, the server may first encode the training article identifier by using One Hot encoding to obtain a training article encoding vector; and then inputting the training article code vector into an embedding layer, and performing densification representation on the training article code vector through the embedding layer to obtain a training article identifier embedding vector. Wherein, the embedded vector of the training article identifier is colidID embedding.

As an example, taking an article sample as a cold article for example, the server vectorizes each training attribute information of the cold article through the embedding layer to obtain a training attribute embedding vector of the training attribute information, and the training attribute embedding vector is recorded as a vector

Embedding training attributes into vectors

The input element stretching network can output a training stretching vector

Wherein, in the step (A),

(ii) a The training stretching vector is used as a stretching function to perform dot multiplication on the vectorized training article identifier of the cold article, so as to obtain an initial war ID embedding (i.e. the training cross feature vector) of the cold article, wherein an expression of the initial war ID embedding is as follows:

then, the server vectorizes the training object identifications of the object samples interacted with the cold goods through the embedding layer to obtain the training object identification embedding vectors of the object samples, and the training object identification embedding vectors are marked as vectors

Embedding training object identification into vector notation

By inputting a meta-bias network, i.e. outputting a training bias vector

Wherein, in the step (A),

(ii) a And fusing the initial war ID embedding of the cold goods by taking the training offset vector as an offset function to obtain final war ID embedding, wherein the expression of the war ID embedding is as follows:

in one embodiment, the server may divide the article sample into a plurality of parts, divide the corresponding training article identifier into a plurality of parts, and then perform vectorization processing on each training article identifier through the embedding layer to obtain a training article identifier embedding vector corresponding to each training article identifier, that is, one training article identifier corresponds to one training article identifier embedding vector. For example, assuming that the article sample is divided into three parts, i.e., a, b, and c, the corresponding training article identifiers also include three parts, i.e., the a-th training article identifier, the b-th training article identifier, and the c-th training article identifier, and accordingly, the training article identifier embedding vector corresponding to the a-th training article identifier can be obtained, and so on.

Correspondingly, the training attribute information and the training object identification of the object sample are also divided into multiple parts correspondingly. And then, performing feature extraction on the training attribute information of each part by using a meta-stretch network to be trained to obtain a training stretch vector of each part, and so on to obtain a training offset vector of each part. For example, assuming that the article sample is divided into three parts, i.e., a, b, and c, the corresponding training attribute information includes three parts, i.e., the a-th training attribute information, the b-th training attribute information, and the c-th training attribute information, and accordingly, the training stretch vector and the training offset vector corresponding to the a-th training attribute information can be obtained.

And then, the server respectively performs feature crossing on each training article identifier and the corresponding training stretching vector, and fuses the obtained training crossing feature vector and the corresponding training offset vector to obtain a training fusion feature vector of each training. For example, assuming that the article sample is divided into three parts, a, b and c, and accordingly three training fused feature vectors, a, b and c, can be obtained.

S408, calculating a first loss value between the training fused feature vector and the sample label.

In one embodiment, when the sample of the object is divided into a plurality of shares, a first loss value between the training fused feature vector and the sample label of each share may be calculated in turn. Wherein, can adopt

The loss function calculates a first loss value. Wherein the content of the first and second substances,

it may be indicated that the fused feature vector is trained,

a sample label may be represented. In addition, other loss functions may be used, and are not particularly limited herein.

And S410, respectively updating network parameters in the meta stretch network and the meta offset network according to the first loss value.

In one embodiment, the server inputs the training fusion feature vector, the vectorized training attribute information and the vectorized training object information of the object sample into the recommendation model, and processes the training fusion feature vector, the vectorized training attribute information and the vectorized training object information of the object sample through the recommendation model to obtain the training recommendation probability. Correspondingly, S410 may specifically include: and when the training recommendation probability meets the training condition, the server respectively updates the network parameters in the meta stretch network and the meta offset network according to the first loss value.

The training object information may refer to object information of the object sample, including training object identification and training object feature information of the object sample. The training object identification can be the identity information or user name of the object sample; the training object feature information includes but is not limited to: the age, sex, place of departure, place of work, school calendar, income, home background, etc. of the target subject. The training recommendation probability is the recommendation probability predicted by the recommendation model in the training process.

In one embodiment, the server obtains training object information of the object sample, inputs the training object information to the embedding layer, and performs vectorization processing on the training object information through the embedding layer to obtain a training object sample embedding vector. The training object sample embedded vector is vectorized training object information.

In One embodiment, the server may first encode training object information of an object sample by using One Hot encoding to obtain a training object encoding vector; and then inputting the training object code vector into an embedding layer, and performing densification representation on the training object code vector through the embedding layer to obtain the training object embedding vector.

In one embodiment, when the article sample and the object sample are divided into a plurality of subsamples, the server performs back propagation on the first loss value of each subsample in the meta-stretch network and the meta-migration network in sequence, so as to obtain a gradient value of a network parameter, and updates the network parameter in the meta-stretch network and the meta-migration network according to the gradient value.

In the above embodiment, feature extraction is performed on training attribute information of an article sample through a meta-stretch network to be trained to obtain a training stretch vector; and extracting the feature of the training object mark of the object sample through a meta-bias network to be trained to obtain a training bias vector, performing feature processing on the training object mark of the object sample which is vectorized based on the training stretching vector and the training bias vector to obtain a training fusion feature vector related to the training object mark, continuously updating network parameters in the meta-stretching network and the meta-bias network through a first loss value between the training fusion feature vector and a sample label even if the object sample is a cold-door object, so that the meta-stretching network and the meta-bias network continuously learn the feature vector related to the training object mark and fused with the object attribute and the object mark of the interactive object, and obtaining the mark feature vector of the cold-door object through the trained meta-stretching network and the meta-bias network even if the cold-door object is recommended and the cold-door object is used by only a small number of users, based on the identification feature vector, the attribute information and the object identification of a small number of object interaction objects, accurate recommendation of cold objects can be achieved.

In one embodiment, as shown in fig. 5, the method further comprises:

s502, vectorization processing is carried out on the training article identification through the embedding layer, and a training article identification embedding vector is obtained.

In one embodiment, the server obtains training article identifications of the article samples, inputs the training article identifications to the embedding layer, and conducts vectorization processing on the training article identifications through the embedding layer to obtain training article identification embedding vectors.

In One embodiment, the server may first encode the training article identifier of the article sample by using One Hot encoding to obtain a training article identifier encoding vector; and then inputting the training article identification code vector into an embedding layer, and performing densification representation on the training article identification code vector through the embedding layer to obtain the training article identification embedding vector.

And S504, calculating a second loss value between the training article identification embedding vector and the sample label.

In one embodiment, when the article sample is divided into multiple parts, the server may sequentially calculate a second loss value between the training article identifier embedding vector and the sample label corresponding to each sub-article sample. Wherein, can adopt

may represent a training article identification embedding vector,

S506, inputting the training fusion feature vector, the vectorized training attribute information and the vectorized training object information of the object sample into a recommendation model, and processing the training fusion feature vector, the vectorized training attribute information and the vectorized training object information of the object sample through the recommendation model to obtain the training recommendation probability.

In one embodiment, the server obtains training object information of the object sample, inputs the training object information to the embedding layer, and performs vectorization processing on the training object information through the embedding layer to obtain a training object embedding vector.

In One embodiment, the server may first encode the training object information in a One Hot encoding manner to obtain a training object encoding vector; and then inputting the training object code vector into an embedding layer, and performing densification representation on the training object code vector through the embedding layer to obtain the training object embedding vector. The training object embedded vector is vectorized training object information.

And S508, when the training recommendation probability meets the training condition, updating the network parameters in the embedded layer according to the second loss value.

In one embodiment, when the article sample and the object sample are divided into a plurality of subsamples, the server performs back propagation on the second loss value of each subsample in the embedding layer, so as to obtain a gradient value of the network parameter of the embedding layer, and updates the network parameter in the embedding layer according to the gradient value.

In the above embodiment, the training article identifier is vectorized by the embedding layer to obtain a training article identifier embedding vector, and the network parameter in the embedding layer is updated according to the second loss value between the training article identifier embedding vector and the sample tag, so that the embedding layer continuously learns the embedding representation of the training article identifier, and even if the training article is a cold article, the embedding representation for preheating can be learned by the embedding layer, which is beneficial to quickly obtaining the preheated training fusion feature vector.

As shown in fig. 6, in one embodiment, a meta-network processing method is provided, which is described by way of example as being performed by the server 104, and includes the following steps:

s602, obtaining training attribute information of the article sample and a training object identification of the object sample.

S604, extracting features of the training attribute information through a meta-stretch network to be trained to obtain a training stretch vector; and performing feature extraction on the training object identification through a to-be-trained meta-offset network to obtain a training offset vector.

And S606, performing feature crossing on the vectorized training article identification and the training stretching vector of the article sample, and fusing the obtained training cross feature vector and the training offset vector to obtain a training fusion feature vector.

S608, calculating a first loss value between the training fusion feature vector and the sample label.

And S610, respectively updating network parameters in the meta stretch network and the meta offset network according to the first loss value.

In one embodiment, the server may input the training fusion feature vectors, the vectorized training attribute information, and the vectorized training object information of the object samples into the recommendation model; processing training fusion characteristic vectors, vectorized training attribute information and vectorized training object information of object samples through a recommendation model to obtain training recommendation probability; and when the training recommendation probability meets the training condition, respectively updating the network parameters in the meta stretch network and the meta offset network according to the first loss value.

In one embodiment, the server may perform vectorization processing on the training object information through the embedding layer to obtain an object sample embedding vector; embedding the object sample into vector to obtain vectorized training object information; calculating a second loss value between the object sample embedding vector and the sample label; and when the training recommendation probability meets the training condition, updating the network parameters in the embedded layer according to the second loss value.

The meta-network training steps described above may refer to S402-S410, and S502-508.

As an example, the item recommendation method may be applied to item cold start, and the effect of item cold start may be increased without affecting the original recommendation system. The item recommendation method can be used in application scenarios such as video platforms, e-commerce platforms, social advertising platforms or other special recommendation systems. The article recommendation method achieves good effects when tests are carried out on video recommendation, e-commerce recommendation and e-commerce display advertisement recommendation.

In practical application, a dedicated recommendation model exists in the application platform or system, in this embodiment, without any change to the recommendation model, two meta networks (meta networks) are learned, that is, a meta stretch network (meta scaling network) and a meta shift network (meta shifting network), when recommending a cold-start item, a stretch function (scaling function) and a shift function (shifting function) are predicted respectively through the two meta networks, then a cold item identification embedding vector (cold ID embedding) is transformed by using the scaling function and the shifting function, and finally, prediction is performed by using the transformed obtained cold ID embedding.

As shown in fig. 7, the overall frame structure is as shown in fig. 7, and measures scaling networks are input to measure scaling networks for the item features or attribute information (item features) of the cold-finished items, and scaling functions are predicted to be brought out. It should be noted that, for items with similar characteristics or attributes, the mapping between cold ID embedding and the rom ID embedding may also be close, and the meta scaling network is defined as follows:

wherein, in the step (A),

。

inputting the ID entries of Users (Interacted Users) who interact with the cold goods into a meta-shifting network for prediction to obtain a shifting function, and further strengthening the output of the scaling function through the shifting function, so that the ID entries are more stable, and the noise influence is reduced (the shifting function is obtained from global interactive prediction). The meta shifting network is defined as follows:

wherein, in the step (A),

。

a war ID Embedding can be obtained through the scaling function and the shifting function, the war ID Embedding can better fit a recommendation model, and the war ID Embedding is expressed as follows:

wherein the content of the first and second substances,

representing an articleiCold ID embedding.

It should be noted that the meta scaling network and the meta shifting network are trained, and the scaling function and the shifting function are predicted by the two meta networks. To train these two meta networks, the training process uses old items to simulate the process of new items (i.e., cold items that currently need to be recommended) joining. Each old items has a cold start stage at the initial stage of adding the platform, and the two meta networks are trained by using the data of the cold start stage of the old items. By learning on similar cold start tasks, the two meta networks can respectively learn to output a function, and cold start cold ID embedding is converted into arm ID embedding through the two functions. The overall training process is shown in table 1:

TABLE 1 training flow sheet

Through the above-mentioned embodiment, can have following technological effect:

the method can be used for cold starting of the articles in various scenes, and the effect of recommending the cold-started articles is improved under the condition that the effect of recommending other articles is not influenced. The model of the original recommendation system is not changed completely, and accurate recommendation of cold goods can be achieved only by additionally training two meta networks.

It should be understood that although the various steps in the flowcharts of fig. 2, 4-6 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, 4-6 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 or stages.

In one embodiment, as shown in fig. 8, there is provided an article recommendation apparatus, 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: an obtaining module 802, a feature extracting module 804, a feature processing module 806, a determining module 808, and a recommending module 810, wherein:

an obtaining module 802, configured to obtain attribute information of an article and an object identifier of an article interaction object;

a feature extraction module 804, configured to perform feature extraction on the attribute information through a meta-stretch network to obtain a stretch vector; extracting the characteristics of the object identification through a meta-offset network to obtain an offset vector;

a feature processing module 806, configured to perform feature crossing on the vectorized article identifier of the article and the stretching vector, and fuse the obtained crossing feature vector with the offset vector to obtain a fused feature vector;

a determining module 808, configured to determine a recommendation probability based on the fused feature vector, the vectorized attribute information, and the vectorized object information of the target object;

and the recommending module 810 is used for recommending the item to the target object when the recommending probability meets a recommending condition.

In one embodiment, as shown in fig. 9, the apparatus further comprises: a vectorization processing module 812; wherein:

a vectorization processing module 812, configured to perform vectorization processing on the attribute information through the embedding layer to obtain an attribute embedded vector;

the feature extraction module 804 is further configured to input the attribute embedding vector to a meta-stretch network; and processing the attribute embedded vector through the meta-stretch network to obtain a stretch vector.

In an embodiment, the vectorization processing module 812 is further configured to perform vectorization processing on the article identifier through the embedding layer to obtain an article identifier embedding vector;

the feature processing module 806 is further configured to perform feature crossing on the item identifier embedding vector and the stretching vector.

In one embodiment, the obtaining module 802 is further configured to obtain object information of the target object;

the vectorization processing module 812 is further configured to perform vectorization processing on the object information through the embedding layer to obtain an object embedding vector;

the feature processing module 806 is further configured to determine a recommendation probability based on the fused feature vector, the attribute embedding vector, and the object embedding vector.

In one embodiment, the object information includes a target object identification and target object characteristic information; the vectorization processing module 812 is further configured to perform vectorization processing on the target object identifier through the embedding layer to obtain a target object identifier embedding vector; vectorizing the target object characteristic information through the embedding layer to obtain a target object characteristic embedding vector; wherein the object embedding vector comprises the target object identification embedding vector and the target object feature embedding vector.

In an embodiment, the determining module 808 is further configured to calculate, through a recommendation model, the fusion feature vector, the attribute embedding vector, the target object identifier embedding vector, and the target object feature embedding vector to obtain a recommendation probability.

In an embodiment, the vectorization processing module 812 is further configured to perform vectorization processing on the object identifier through the embedding layer to obtain an object identifier embedding vector;

a feature extraction module 804, further configured to input the object identifier embedding vector to a meta-migration network; and processing the object identification embedded vector through the meta-migration network to obtain a migration vector.

In one embodiment, the meta stretch network and the meta bias network are obtained by performing meta network processing on the meta stretch network and the meta bias network to be trained, respectively; as shown in fig. 9, the apparatus further includes: a calculation module 814 and an update module 816; wherein:

the obtaining module 802 is further configured to obtain training attribute information of the article sample and a training object identifier of the object sample;

the feature extraction module 804 is further configured to perform feature extraction on the training attribute information through a meta stretch network to be trained to obtain a training stretch vector; performing feature extraction on the training object identification through a to-be-trained meta-offset network to obtain a training offset vector;

the feature processing module 806 is further configured to perform feature crossing on the vectorized training article identifier of the article sample and the training stretching vector, and fuse the obtained training crossing feature vector with the training offset vector to obtain a training fused feature vector;

a calculating module 814, configured to calculate a first loss value between the training fused feature vector and a sample label;

an updating module 816, configured to update network parameters in the meta stretch network and the meta-offset network respectively according to the first loss value.

In one embodiment, the determining module 808 is further configured to input the training fusion feature vector, the vectorized training attribute information, and the vectorized training object information of the object sample into a recommendation model; processing the training fusion feature vector, the vectorized training attribute information and the vectorized training object information of the object sample through the recommendation model to obtain a training recommendation probability;

the updating module 816 is further configured to update network parameters in the meta stretch network and the meta offset network respectively according to the first loss value when the training recommendation probability satisfies a training condition.

In an embodiment, the vectorization processing module 812 is further configured to perform vectorization processing on the training article identifier through the embedding layer to obtain a training article identifier embedding vector;

a calculating module 814, further configured to calculate a second loss value between the training item identifier embedding vector and the sample label;

the updating module 816 is further configured to update the network parameter in the embedded layer according to the second loss value when the training recommendation probability meets the training condition.

For the specific definition of the item recommendation device, reference may be made to the above definition of the item recommendation method, which is not described herein again. The modules in the article recommendation device can be wholly or partially implemented by software, hardware and a combination 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, as shown in fig. 10, a meta-network processing apparatus 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: an obtaining module 1002, a feature extracting module 1004, a feature processing module 1006, a calculating module 1008, and an updating module 1010, wherein:

an obtaining module 1002, configured to obtain training attribute information of an article sample and a training object identifier of an object sample;

a feature extraction module 1004, configured to perform feature extraction on the training attribute information through a meta stretch network to be trained, to obtain a training stretch vector; performing feature extraction on the training object identification through a to-be-trained meta-offset network to obtain a training offset vector;

a feature processing module 1006, configured to perform feature crossing on the vectorized training article identifier of the article sample and the training stretching vector, and fuse the obtained training cross feature vector with the training offset vector to obtain a training fusion feature vector;

a calculating module 1008, configured to calculate a first loss value between the training fusion feature vector and a sample label;

an updating module 1010, configured to update network parameters in the meta stretch network and the meta-bias network according to the first loss value.

In one embodiment, as shown in fig. 11, the apparatus further comprises: a determination module 1012; wherein:

a determining module 1012, configured to input the training fusion feature vector, the vectorized training attribute information, and the vectorized training object information of the object sample into a recommendation model; processing the training fusion feature vector, the vectorized training attribute information and the vectorized training object information of the object sample through the recommendation model to obtain a training recommendation probability;

the updating module 1010 is further configured to update network parameters in the meta stretch network and the meta offset network respectively according to the first loss value when the training recommendation probability satisfies a training condition.

In one embodiment, as shown in fig. 11, the apparatus further comprises: a vectorization processing module 1014; wherein:

the vectorization processing module 1014 is configured to perform vectorization processing on the training article identifier through the embedding layer to obtain a training article identifier embedding vector;

a calculating module 1008, further configured to calculate a second loss value between the training item identifier embedding vector and the sample label;

the updating module 1010 is further configured to update the network parameters in the embedded layer according to the second loss value when the training recommendation probability satisfies a training condition.

For the specific definition of the meta-network processing device, reference may be made to the above definition of the meta-network processing method, which is not described herein again. The respective modules in the above-described meta-network processing apparatus may be wholly or partially implemented by software, hardware, and a combination 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. 12. 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 device is used for storing information of the article and information of the corresponding object. 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 item recommendation method or a meta-network processing method.

Those skilled in the art will appreciate that the architecture shown in fig. 12 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 item recommendation method, characterized in that the method comprises:

extracting the characteristics of the attribute information through a meta-stretch network to obtain a stretch vector; extracting the characteristics of the object identification through a meta-offset network to obtain an offset vector; the meta-stretch network is a meta-network which is formed based on a stretch parameter and a variable of an attribute embedding vector representing the article and is used for extracting the feature of the attribute information, and the meta-offset network is a meta-network which is formed based on an offset parameter and a variable of an object identification embedding vector representing the article interaction object and is used for extracting the feature of the object identification;

2. The method of claim 1, further comprising:

vectorizing the attribute information through an embedding layer to obtain an attribute embedding vector;

the extracting the features of the attribute information through the meta-stretch network to obtain a stretch vector includes:

inputting the attribute embedding vector to a meta-stretch network;

and processing the attribute embedded vector through the meta-stretch network to obtain a stretch vector.

3. The method of claim 2, further comprising:

vectorizing the article identification through the embedding layer to obtain an article identification embedding vector;

said feature interleaving the vectorized item identification of the item with the stretch vector comprises:

performing feature intersection on the article identification embedding vector and the stretching vector.

4. A method according to claim 2 or 3, characterized in that the method further comprises:

acquiring object information of a target object;

vectorizing the object information through the embedding layer to obtain an object embedding vector;

the determining a recommendation probability based on the fused feature vector, the vectorized attribute information, and the vectorized object information of the target object comprises:

determining a recommendation probability based on the fused feature vector, the attribute embedding vector, and the object embedding vector.

5. The method of claim 4, wherein the object information comprises a target object identification and target object characteristic information; the vectorizing the object information through the embedding layer to obtain an object embedding vector includes:

vectorizing the target object identifier through the embedding layer to obtain a target object identifier embedding vector; vectorizing the target object characteristic information through the embedding layer to obtain a target object characteristic embedding vector;

wherein the object embedding vector comprises the target object identification embedding vector and the target object feature embedding vector.

6. The method of claim 5, wherein determining a recommendation probability based on the fused feature vector, the attribute embedding vector, and the object embedding vector comprises:

and calculating the fusion feature vector, the attribute embedding vector, the target object identification embedding vector and the target object feature embedding vector through a recommendation model to obtain recommendation probability.

7. The method according to any one of claims 1 to 3, further comprising:

vectorizing the object identifier through an embedding layer to obtain an object identifier embedding vector;

the extracting the features of the object identifier through the meta-migration network to obtain the migration vector comprises:

inputting the object identification embedding vector to a metamigration network;

and processing the object identification embedded vector through the meta-migration network to obtain a migration vector.

8. The method according to any one of claims 1 to 3, wherein the meta stretch network and the meta bias network are obtained by performing meta network processing on the meta stretch network and the meta bias network to be trained, respectively; the meta network processing includes:

9. The method of claim 8, further comprising:

10. The method of claim 9, further comprising:

11. A meta-network processing method, the method comprising:

extracting the characteristics of the training attribute information through a meta-stretch network to be trained to obtain a training stretch vector; performing feature extraction on the training object identification through a to-be-trained meta-offset network to obtain a training offset vector; the meta stretch network is a meta network which is formed based on a stretch parameter and a variable of a training attribute embedding vector representing the article sample and is used for extracting the feature of the training attribute information, and the meta offset network is a meta network which is formed based on an offset parameter and a variable of a training object identification embedding vector representing the object sample and is used for extracting the feature of the training object identification;

12. The method of claim 11, further comprising:

13. The method of claim 12, further comprising:

14. An item recommendation device, the device comprising:

the characteristic extraction module is used for extracting the characteristics of the attribute information through a meta-stretch network to obtain a stretch vector; extracting the characteristics of the object identification through a meta-offset network to obtain an offset vector; the meta-stretch network is a meta-network which is formed based on a stretch parameter and a variable of an attribute embedding vector representing the article and is used for extracting the feature of the attribute information, and the meta-offset network is a meta-network which is formed based on an offset parameter and a variable of an object identification embedding vector representing the article interaction object and is used for extracting the feature of the object identification;

15. The apparatus of claim 14, further comprising:

the vectorization processing module is used for carrying out vectorization processing on the attribute information through the embedding layer to obtain an attribute embedding vector;

the feature extraction module is further used for inputting the attribute embedding vector to a meta-stretch network; and processing the attribute embedded vector through the meta-stretch network to obtain a stretch vector.

16. The apparatus according to claim 15, wherein the vectorization processing module is further configured to perform vectorization processing on the article identifier through the embedding layer to obtain an article identifier embedding vector;

the feature processing module is further configured to perform feature crossing on the article identifier embedding vector and the stretching vector.

17. The apparatus according to claim 15 or 16, wherein the obtaining module is further configured to obtain object information of a target object;

the vectorization processing module is further configured to perform vectorization processing on the object information through the embedding layer to obtain an object embedding vector;

the determination module is further configured to determine a recommendation probability based on the fused feature vector, the attribute embedding vector, and the object embedding vector.

18. The apparatus of claim 17, wherein the object information comprises a target object identification and target object characteristic information; the vectorization processing module is further configured to perform vectorization processing on the target object identifier through the embedding layer to obtain a target object identifier embedding vector; vectorizing the target object characteristic information through the embedding layer to obtain a target object characteristic embedding vector;

19. The apparatus of claim 18, wherein the determining module is further configured to compute the fusion feature vector, the attribute embedding vector, the target object id embedding vector, and the target object feature embedding vector through a recommendation model to obtain a recommendation probability.

20. The apparatus of any one of claims 14 to 16, further comprising:

the vectorization processing module is used for carrying out vectorization processing on the object identifier through the embedding layer to obtain an object identifier embedding vector;

the feature extraction module is further configured to input the object identifier embedding vector to a meta-migration network; and processing the object identification embedded vector through the meta-migration network to obtain a migration vector.

21. The apparatus according to any one of claims 15 to 16, wherein the meta stretch network and the meta bias network are obtained by performing meta network processing on the meta stretch network and the meta bias network to be trained, respectively; the device further comprises:

the acquisition module is further used for acquiring training attribute information of the article sample and a training object identifier of the object sample;

the feature extraction module is further configured to perform feature extraction on the training attribute information through a meta stretch network to be trained to obtain a training stretch vector; performing feature extraction on the training object identification through a to-be-trained meta-offset network to obtain a training offset vector;

the feature processing module is further configured to perform feature crossing on the vectorized training article identifier of the article sample and the training stretching vector, and fuse the obtained training crossing feature vector and the training offset vector to obtain a training fused feature vector;

22. The apparatus of claim 21, wherein the determining module is further configured to input the training fusion feature vector, the vectorized training attribute information, and the vectorized training object information for the object sample into a recommendation model; processing the training fusion feature vector, the vectorized training attribute information and the vectorized training object information of the object sample through the recommendation model to obtain a training recommendation probability;

and the updating module is further configured to update the network parameters in the meta stretch network and the meta offset network respectively according to the first loss value when the training recommendation probability satisfies a training condition.

23. The apparatus of claim 22, wherein the vectorization processing module is further configured to perform vectorization processing on the training article identifier through an embedding layer to obtain a training article identifier embedding vector;

the calculation module is further configured to calculate a second loss value between the training article identifier embedding vector and the sample label;

and the updating module is further used for updating the network parameters in the embedded layer according to the second loss value when the training recommendation probability meets the training condition.

24. A meta-network processing apparatus, the apparatus comprising:

the feature extraction module is used for extracting features of the training attribute information through a meta-stretch network to be trained to obtain a training stretch vector; performing feature extraction on the training object identification through a to-be-trained meta-offset network to obtain a training offset vector; the meta stretch network is a meta network which is formed based on a stretch parameter and a variable of a training attribute embedding vector representing the article sample and is used for extracting the feature of the training attribute information, and the meta offset network is a meta network which is formed based on an offset parameter and a variable of a training object identification embedding vector representing the object sample and is used for extracting the feature of the training object identification;

25. The apparatus of claim 24, further comprising:

a determining module, configured to input the training fusion feature vector, the vectorized training attribute information, and the vectorized training object information of the object sample into a recommendation model; processing the training fusion feature vector, the vectorized training attribute information and the vectorized training object information of the object sample through the recommendation model to obtain a training recommendation probability;

26. The apparatus of claim 25, further comprising:

the vectorization processing module is used for carrying out vectorization processing on the training article identification through the embedding layer to obtain a training article identification embedding vector;

27. 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 13 when executing the computer program.

28. 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 13.