CN116205700A

CN116205700A - Recommendation method and device for target product, computer equipment and storage medium

Info

Publication number: CN116205700A
Application number: CN202111450911.0A
Authority: CN
Inventors: 石志林
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2021-11-30
Filing date: 2021-11-30
Publication date: 2023-06-02

Abstract

The application relates to a recommendation method and device of target products, computer equipment and storage media. The method can be applied to a product pushing scene of the vehicle-mounted terminal, and comprises the following steps: acquiring product recommendation characteristics of a target product, object characteristics of candidate objects and behavior characteristic sequences obtained by product interaction of the candidate objects under each product interaction scene; determining the interest feature of the candidate object on the target product based on the behavior feature in the behavior feature sequence and the product recommendation feature; determining the interest score of the candidate object to the target product according to the product recommendation feature, the object feature and the splicing feature between the interest features; selecting a target object with the interest score reaching a scoring condition from the candidate objects; and pushing the product recommendation information of the target product to the target object. By adopting the method, the accuracy of product recommendation can be improved.

Description

Recommendation method and device for target product, computer equipment and storage medium

Technical Field

The present disclosure relates to the field of internet technologies, and in particular, to a method and apparatus for recommending a target product, a computer device, and a storage medium.

Background

With the development of internet technology, information acquisition through interconnection becomes a part of life, entertainment and work of people. Merchants often recommend products via the internet in order to increase awareness or sales.

In the existing product recommendation method, the interests of an object are mined according to the behavior data of the object in a single scene, and modeling is performed based on the mined interests to recommend products. However, the interests of the object are often diverse, and the preference of the object cannot be expressed truly and comprehensively based on the interests of a single scene, so that the accuracy of product recommendation is low.

Disclosure of Invention

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

A method of recommending a target product, the method comprising:

acquiring product recommendation characteristics of a target product, object characteristics of candidate objects and behavior characteristic sequences obtained by product interaction of the candidate objects under each product interaction scene;

determining the interest feature of the candidate object on the target product based on the behavior feature in the behavior feature sequence and the product recommendation feature;

Determining the interest score of the candidate object to the target product according to the product recommendation feature, the object feature and the splicing feature between the interest features;

selecting a target object with the interest score reaching a scoring condition from the candidate objects;

and pushing the product recommendation information of the target product to the target object.

A recommendation device for a target product, the device comprising:

the feature acquisition module is used for acquiring product recommendation features of a target product, object features of candidate objects and behavior feature sequences obtained by product interaction of the candidate objects under each product interaction scene;

the interest feature determining module is used for determining the interest feature of the candidate object on the target product based on the behavior feature in the behavior feature sequence and the product recommendation feature;

the interest score determining module is used for determining the interest score of the candidate object to the target product according to the product recommendation characteristic, the object characteristic and the splicing characteristic among the interest characteristics;

the object selection module is used for selecting a target object with the interest score reaching a score condition from the candidate objects;

And the product recommendation module is used for pushing the product recommendation information of the target product to the target object.

In one embodiment, the apparatus further comprises:

the behavior information ordering module is used for ordering the behavior information in the behavior information set according to the behavior time;

the feature acquisition module is further configured to:

and carrying out vectorization processing on the product recommendation information, the object information and the behavior information in the ordered behavior information set.

In one embodiment, the interest feature determination module is further configured to:

acquiring behavior time and position codes corresponding to behavior features in the behavior feature sequence;

determining sub-interest features corresponding to the behavior features based on the behavior features, the behavior time, the position codes and the product recommendation features in the behavior feature sequence;

and determining the interest characteristic of the candidate object to the target product according to each sub-interest characteristic.

extracting the characteristics of the behavior characteristic sequence based on a multi-head attention mechanism to obtain corresponding attention behavior characteristics; the number of the attention behavior features corresponding to each behavior feature is consistent with the number of heads of the multi-head attention mechanism;

Aiming at each attention behavior feature, determining a sub-interest feature corresponding to the current attention behavior feature based on the current attention behavior feature, the product recommendation feature of the target product, the behavior time and the position code until the sub-interest feature corresponding to each attention behavior feature is obtained;

and splicing sub-interest features of the attention behavior features corresponding to each behavior feature to obtain sub-interest features corresponding to each behavior feature.

In one embodiment, the feature acquisition module is further configured to:

acquiring context information formed by product interaction of the candidate object under each product interaction scene;

carrying out vectorization processing and dimension reduction processing on the context information in sequence to obtain context characteristics;

the interest score determination module is further configured to:

splicing the product recommended feature, the object feature, the interest feature and the context feature to obtain a spliced feature;

and determining the interest score of the candidate object to the target product based on the splicing characteristics.

In one embodiment, the product recommendation feature, the object feature, the behavioral feature sequence, and the contextual feature are feature processed by a product recommendation model; the apparatus further comprises:

The training feature acquisition module is used for inputting training product recommendation information corresponding to a sample product, training object information corresponding to a sample object and training behavior information sets obtained by the interaction of the sample object on the sample product in each product interaction scene into the product recommendation model to be trained to obtain training product recommendation features, training object features and training behavior feature sequences;

the training interest feature acquisition module is used for determining training interest features of the sample object on the sample product based on the behavior features in the training behavior feature sequence and the training product recommendation features;

the training interest score determining module is used for determining the interest score of the sample object to the sample product according to the recommended feature of the training product, the characteristic of the training object and the splicing characteristic among the training interest features;

a loss value determining module, configured to determine a first loss value between the interest score and an interest tag;

and the model parameter adjustment module is used for adjusting parameters of the product recommendation model to be trained based on the first loss value.

In one embodiment, the loss value determination module is further configured to:

Based on the behavior characteristics in the training behavior characteristic sequence, determining the predicted behavior characteristics corresponding to the sample object at the next moment;

determining a second loss value based on the predicted behavioral characteristics and behavioral characteristic signatures;

the model parameter adjustment module is further configured to:

and carrying out parameter adjustment on the product recommendation model to be trained based on the first loss value and the second loss value.

In one embodiment, the model parameter adjustment module is further configured to:

obtaining a loss coefficient corresponding to the second loss value;

determining a model loss value from the loss coefficient, the first loss value, and the second loss value;

and carrying out parameter adjustment on the product recommendation model to be trained based on the model loss value, and stopping training until the obtained model loss value reaches a preset condition.

In one embodiment, the training feature acquisition module is further configured to:

acquiring training context information formed by the interaction of the sample object with the sample product under each product interaction scene;

inputting the training context information into the product recommendation model to be trained to obtain training context characteristics;

The training interest score determining module is further configured to:

and determining the interest scores of the sample objects on the sample products according to the recommended features of the training products, the features of the training objects, the spliced features between the training interest features and the training context features.

A computer device comprising a memory storing a computer program and a processor which when executing the computer program performs the steps of:

A computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of:

A computer program comprising computer instructions stored in a computer readable storage medium, the computer instructions being read from the computer readable storage medium by a processor of a computer device, the processor executing the computer instructions causing the computer device to perform the steps of:

According to the recommendation method, the recommendation device, the computer equipment and the storage medium of the target product, the product recommendation characteristics of the target product, the object characteristics of the candidate object and the behavior characteristic sequences obtained by product interaction of the candidate object under each product interaction scene are obtained; the interest characteristics of the candidate object to the target product are determined based on the behavior characteristics and the product recommendation characteristics in the behavior characteristic sequence, so that the interest of the candidate object in each scene can be comprehensively considered to obtain the real preference of the candidate object, the interest score of the candidate object to the target product is determined according to the product recommendation characteristics, the object characteristics and the splicing characteristics among the interest characteristics, the target object with the interest score reaching the score condition is selected from the candidate object, and the accuracy of product recommendation is improved when the product recommendation information of the target product is pushed to the target object.

Drawings

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

FIG. 2 is a flowchart of a recommendation method for a target product according to an embodiment;

FIG. 3 is a flow chart of a product recommendation model training step in one embodiment;

FIG. 4 is a flowchart of a recommendation method for a target product according to another embodiment;

FIG. 5 is a flowchart of a recommendation method for a target product according to another embodiment;

FIG. 6 is a schematic diagram of a product recommendation model in one embodiment;

FIG. 7 is a block diagram of a recommendation device for a target product in one embodiment;

FIG. 8 is a block diagram illustrating a recommendation device for a target product according to another embodiment;

FIG. 9 is an internal block diagram of a computer device in one embodiment;

fig. 10 is an internal structural view of a computer device in another embodiment.

Detailed Description

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

The recommendation method of the target product provided by the application relates to the technologies of artificial intelligence such as machine learning, natural language processing and the like, wherein:

Artificial intelligence (Artificial Intelligence, AI) is the theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and extend human intelligence, sense the environment, acquire knowledge and use the knowledge to obtain optimal results. In other words, artificial intelligence is an integrated technology of computer science that attempts to understand the essence of intelligence and to produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence, i.e. research on design principles and implementation methods of various intelligent machines, enables the machines to have functions of sensing, reasoning and decision.

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

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

Natural language processing (Nature Language processing, NLP) is an important direction in the fields of computer science and artificial intelligence. It is studying various theories and methods that enable effective communication between a person and a computer in natural language. Natural language processing is a science that integrates linguistics, computer science, and mathematics. Thus, the research in this field will involve natural language, i.e. language that people use daily, so it has a close relationship with the research in linguistics. Natural language processing techniques typically include text processing, semantic understanding, machine translation, robotic questions and answers, knowledge graph techniques, and the like.

With research and progress of artificial intelligence technology, research and application of artificial intelligence technology are being developed in various fields, such as common smart home, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned, autopilot, unmanned, robotic, smart medical, smart customer service, car networking, autopilot, smart transportation, etc., and it is believed that with the development of technology, artificial intelligence technology will be applied in more fields and will be of increasing importance.

The recommendation method of the target product can be applied to a recommendation system of the target product shown in fig. 1. As shown in fig. 1, the recommendation system for the target product includes a terminal 102 and a server 104, where the terminal 102 communicates with the server 104 through a network. In one embodiment, the terminal 102 and the server 104 may each independently execute the recommendation method of the target product provided in the embodiments of the present application. The terminal 102 and the server 104 may also cooperate to perform the recommendation method for the target product provided in the embodiments of the present application. When the server 104 cooperates with the recommendation method for executing the target product provided in the embodiment of the present application, the server acquires the product recommendation feature of the target product, the object feature of the candidate object, and the behavior feature sequence obtained by product interaction of the candidate object under each product interaction scene; determining the interest feature of the candidate object on the target product based on the behavior feature and the product recommendation feature in the behavior feature sequence; determining the interest score of the candidate object to the target product according to the product recommendation feature, the object feature and the splicing feature between the interest features; selecting a target object with interest score reaching a score condition from the candidate objects; and pushing product recommendation information of the target product to the target object.

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

Wherein the terminal 102 communicates with the server 104 via a network, such as a wired or wireless network. The terminal 102 may be, but not limited To, various desktop computers, notebook computers, smart phones, tablet computers, intelligent voice interaction devices, intelligent home appliances, vehicle terminals, portable wearable devices, etc., and the server 104 may be an independent physical server or may be a service node in a blockchain system, where a Peer-To-Peer (P2P, peer To Peer) network is formed between service nodes in the blockchain system, and the P2P protocol is an application layer protocol running on top of a transmission control protocol (TCP, transmission Control Protocol) protocol. In addition, the server 104 may also be a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (Content Delivery Network, CDN), and basic cloud computing services such as big data and artificial intelligence platforms.

In one embodiment, as shown in fig. 2, a recommendation method for a target product is provided, and the method is applied to the server in fig. 1 for illustration, and includes the following steps:

s202, obtaining product recommendation characteristics of a target product, object characteristics of candidate objects and behavior characteristic sequences obtained by product interaction of the candidate objects under each product interaction scene.

The target product is a product which can be used for being pushed to an object, and specifically can be at least one of various information and various articles. The various types of information include, but are not limited to, at least one of an application, text, expression, picture, audio, video, file, link, etc. Various types of items may include physical items and virtual items. The physical articles comprise various physical products, in particular various electronic products such as mobile phones, computers, notebooks, watches, and the like, and also apparel products such as clothes, shoes, and the like, and the physical articles are not limited excessively.

Virtual items include, but are not limited to, insurance products, financial products, virtual gift resources, virtual scenes, virtual characters, virtual props, and the like. The virtual scene can be a game scene, a virtual reality simulation scene and the like in the game equipment, the virtual characters can be various characters in the game, and the virtual props can be various props and the like in the game.

The candidate objects are objects capable of recommending target products, the product interaction scene refers to a scene corresponding to the product when the product interaction scene generates interaction behaviors, the product interaction scene can be an online shopping scene, an application program browsing scene, a video playing scene, a social chat scene and the like, the product is interacted, namely the interaction behaviors with the product are generated, and the interaction behaviors comprise clicking, accessing, purchasing, downloading, playing, collecting and the like. For example, when the product interaction scene is an online shopping scene, the product is an article serving as a commodity, the candidate object can purchase the product, so as to generate corresponding behavior information, the generated behavior information includes product recommendation information, interaction behavior, time and the like corresponding to the interacted product, and the time can be time when the interaction behavior occurs and/or time difference between the time when the interaction behavior occurs and the current time.

The product recommendation characteristics of the target product are obtained by carrying out characteristic processing on product recommendation information of the target product; the object characteristics of the candidate objects are obtained by carrying out characterization processing on object information of the candidate objects; the behavior feature sequence is obtained by carrying out characteristic processing on a behavior information set, and particularly is a sequence obtained by sequencing a plurality of behavior features according to the occurrence time of interaction behaviors.

The product recommendation information refers to attribute information of a product, and comprises information such as a product name, a product number, a product category, product semantics, a product image and the like; the object information comprises information such as sex, age, academic and regional of the object; the behavior information comprises product recommendation information, interaction behavior, time and the like corresponding to the interacted products, and the time can be the time of the interaction behavior and/or the time difference between the time of the interaction behavior and the current time.

It is understood that, in the case of obtaining the authorization of each object, the server may collect and store the object information of each object and the behavior information of each object, and acquire the stored object information and the behavior information of each object when the object information of each object and the behavior information of each object need to be processed. Wherein the collection, use and processing of the object information of each object and the behavior information of each object is required to comply with the relevant laws and regulations and standards of the relevant country and region.

Specifically, when the product pushing condition is met, the server acquires product recommendation information of a target product, objects of candidate objects and a behavior information set obtained by product interaction of the candidate objects under each product interaction scene, and performs characterization processing on each acquired information to obtain product recommendation characteristics of the target product, object characteristics of the candidate objects and a behavior characteristic sequence obtained by product interaction of the candidate objects under each product interaction scene.

The product pushing conditions include a time condition and a request condition, wherein the time condition can be specifically a preset product pushing period, for example, the product pushing period is 24 hours, the pushing time is 10 am, and the product pushing condition is satisfied when the ten am of each day is reached; the request condition may specifically be that a product push request is received, and the server may receive a product recommendation request sent from at least one terminal.

The characterization processing includes vectorization processing and dimension reduction processing, wherein vectorization refers to converting information into high-dimension vectors so as to process the information through an artificial intelligent model, and dimension reduction processing refers to reducing redundant information of the information in the form of the high-dimension vectors to obtain low-dimension vectors so as to improve the processing speed of the information.

In one embodiment, S202 specifically includes the steps of: acquiring product recommendation information of a target product, object information of a candidate object and a behavior information set obtained by product interaction of the candidate object under each product interaction scene; vectorizing each piece of behavior information in the product recommendation information, the object information and the behavior information set; and performing dimension reduction treatment on the vectorized result to obtain product recommended features, object features and behavior feature sequences.

Specifically, after obtaining product recommendation information of a target product, object information of a candidate object and a behavior information set of the candidate object, a server carries out vectorization processing on the product recommendation information of the target product, the object information of the candidate object and the behavior information set of the candidate object to obtain a product recommendation information vector, an object information vector and a behavior information vector set in a high-dimensional space, then carries out dimension reduction processing on the product recommendation information vector, the object information vector and the behavior information vector set in the high-dimensional space by adopting a preset dimension reduction algorithm to obtain product recommendation features, object features and a behavior feature set, and determines a behavior feature sequence according to the behavior feature set.

Wherein the dimension reduction algorithm may be at least one of a principal component analysis algorithm (PCA), a singular value decomposition algorithm (SVD), a factor analysis algorithm (FA), an independent component analysis algorithm (ICA), and a linear discriminant analysis algorithm (LDA).

In one embodiment, the process of determining the behavior feature sequence by the server according to the behavior feature set may specifically be that each behavior feature in the behavior feature set is ordered according to the behavior time, so as to obtain the behavior feature sequence.

In one embodiment, the product recommendation information vector, the object information vector, and the behavior information vector in the behavior information vector set may all be represented in one-hot vector form, and the product recommendation feature, the object feature, and the behavior feature in the behavior feature sequence may all be represented as ebedding vectors. Wherein one-hot vector is a sparse vector representation with a value of 0 or 1, one vector has only one 1, and the other values are 0; the Embedding can encode the object with the low-dimensional vector and can also keep the meaning, and the nature of the embdding vector is that the objects corresponding to the vectors with similar distances can have similar meanings.

For example, a product recommendation is characterized by x _u The object features x _v The behavior feature sequence is x _b Wherein x is _b ＝{e ₁ ,e ₂ ,…,e _i ,…,e _n }，e _i For the ith behavioral feature in the behavioral feature sequence, n is the number of behavioral features in the behavioral feature sequence,

for the behavior feature e _i Is (j) th information, e.g.)>

For the behavior feature e _i Corresponding commodity name,/->

For the behavior feature e _i Corresponding commodity ID->

For the behavior feature e _i Corresponding interaction behavior, < >>

For the behavior feature e _i Corresponding time, etc. It will be appreciated that->

Is an Embedding dictionary, < >>

Is an assembled vector of length D.

In one embodiment, the product name, the product number, the product category, the product semantics and the like are vectorized to obtain corresponding text vectors, and the product image is vectorized to obtain corresponding image feature vectors; performing vectorization processing on object information of candidate objects to obtain object information vectors in a high-dimensional space, and performing vectorization processing on gender, age, academic and region of the objects to obtain corresponding text vectors; and carrying out vectorization processing on the interaction behavior, the time of the interaction behavior and the like, and obtaining a corresponding text vector.

In one embodiment, before vectorizing the product recommendation information of the target product, the object information of the candidate object, and the behavior information set of the candidate object, the server may further perform preprocessing on vectorizing the product recommendation information of the target product, the object information of the candidate object, and the behavior information set of the candidate object, where the preprocessing mode includes normalization processing, numerical conversion, and encoding processing, and specifically includes the following steps: determining the product recommendation information of a target product, the object information of a candidate object and the information type of the behavior information of the candidate object, determining a target preprocessing mode based on the information type of the information, preprocessing the product recommendation information, the object information and the behavior information set according to the target preprocessing mode, and obtaining preprocessed product recommendation information, preprocessed object information and preprocessed behavior information set.

The preprocessing mode comprises normalization processing, numerical conversion and encoding processing, the information type comprises continuous type and discrete type, the target preprocessing mode matched with the continuous type is normalization processing or numerical conversion processing, and the target preprocessing mode matched with the discrete type is encoding processing. The continuous information is information that is index-value-type and is continuous in value, for example, information that is continuous, such as age of an object, time of occurrence of interaction, time difference between occurrence of interaction and current time, and discrete information, such as product name, product number, product category, product semantics, product image, gender, academic, region, interaction, and the like.

The normalization processing refers to scaling the values of the continuous information so that the obtained processed values are within a target value range. For example, in the embodiment of the present application, the following formula may be specifically used to normalize the age:

W _age ＝|Age/10| (1)

wherein Age of Age is W _age Pre-processed information corresponding to the age of the subject.

For the time information, the following formula (2) may be used for the time information numerical conversion processing to obtain a conversion result:

value＝e ^-t*0.2 (2)

Wherein t represents time in the behavior information, and value is a conversion result.

The feature coding processing refers to determining codes corresponding to discrete information respectively, and taking the codes as preprocessed information corresponding to the information.

For example, for the gender of the user, the gender "male" may be indicated by the code 1, and the gender "female" may be indicated by the code 0; for regions, a region mapping table can be established, in the region mapping table, beijing is mapped to be "1", shanghai is mapped to be "2", guangzhou is mapped to be "3", shenzhen is mapped to be "4", and the like, so that corresponding numbers are respectively mapped to each city.

In the above embodiment, the server performs vectorization processing and dimension reduction processing on the obtained product recommendation information of the target product, the object information of the candidate object, and the behavior information set obtained by product interaction of the candidate object under each product interaction scene, so that each information can be conveniently processed by an artificial intelligent model, redundant information in the information to be processed is reduced, and the processing speed of each information by the artificial intelligent model is improved.

S204, based on the behavior characteristics and the product recommendation characteristics in the behavior characteristic sequence, the interest characteristics of the candidate object to the target product are determined.

It can be understood that the behavior characteristics of the behavior characteristic sequence carry product recommendation information of the product interacted in the historical time period of the candidate object and the interactive behavior executed on the product, based on the behavior characteristics and the target product recommendation characteristics in the behavior characteristic sequence, the correlation between the product recommendation information of the product interacted in the historical time period and the product recommendation characteristics of the target product can be determined, and further, based on the correlation and the interactive behavior executed on the product interacted in the historical time period of the object, the interest characteristics of the candidate object on the target product can be determined.

In one embodiment, S204 specifically includes the steps of: acquiring behavior time and position codes corresponding to behavior features in a behavior feature sequence; determining sub-interest features corresponding to the behavior features based on the behavior features, the behavior time, the position codes and the product recommendation features in the behavior feature sequence; and determining the interest characteristics of the candidate object on the target product according to the interest characteristics.

Wherein position coding refers to the coding order of the positions of the behavior features in the behavior feature sequence, e.g. behavior feature sequence x _b ＝{e ₁ ,e ₂ ,…,e _i ,…,e _n In the }, behavior feature e _i The position code of (c) can be expressed as p _i Behavior characteristics e _i The behavior time of (c) can be expressed as t _i 。

Specifically, after obtaining the behavior time and the position code corresponding to the behavior feature in the behavior feature sequence, the server determines a time coefficient according to the behavior time, determines sub-interest features corresponding to the behavior features according to the behavior features, the time coefficient, the position code and the product recommendation feature, and determines the interest feature of the candidate object on the target product according to the sub-interest features corresponding to the behavior features.

In one embodiment, the server determines a time difference of the current time of the behavioral time interval according to the behavioral time of the behavioral feature, discretizes the time difference to obtain a time coefficient, and determines a sub-interest feature corresponding to the behavioral feature based on the time coefficient, the behavioral feature, the position code and the product recommendation feature by using the following sub-interest feature determination function. Wherein the expression of the sub-interest feature determination function is as follows:

Interest _i ＝α _i (e _i ,x _u ,p _i )×e _i (3)

wherein, the Intrest _i For the ith behavioral characteristic e _i Corresponding sub-interest feature, alpha _i For the ith behavioral characteristic e _i Corresponding time coefficient, x _u Product recommendation features for target products, p _i For the ith behavioral characteristic e _i The corresponding position is encoded.

In one embodiment, after obtaining sub-interest features corresponding to each behavior feature, the server superimposes each sub-interest feature, so as to obtain an interest feature of the candidate object on the target product, and specifically, the interest feature may be determined by adopting the following formula:

wherein, the Interest is the Interest characteristic of the candidate object to the target product _i I-th behavioral characteristic e of candidate object _i The corresponding sub-interest feature.

In the above embodiment, the server determines the sub-interest feature corresponding to each behavior feature by acquiring the behavior time and the position code corresponding to the behavior feature in the behavior feature sequence, based on the behavior feature, the behavior time, the position code and the product recommendation feature in the behavior feature sequence, and determines the interest feature of the candidate object on the target product according to each sub-interest feature, that is, determines the interest feature of the candidate object on the target product based on the correlation between the product in the behavior feature sequence and the target product, and further can improve the accuracy of product pushing when determining product pushing based on the interest feature.

S206, determining interest scores of candidate objects on target products according to the product recommendation characteristics, the object characteristics and the splicing characteristics among the interest characteristics.

The interest score is used to characterize the interest degree of the candidate object in the target product, which can be said to be the probability of the candidate object interacting with the target product, and it can be understood that the greater the interest degree of the candidate object in the target product, the greater the probability of the candidate object interacting with the target product after pushing the target product to the candidate object, wherein the interaction behavior can be clicking, accessing, purchasing, downloading, playing, collecting and other behaviors.

Specifically, after obtaining the product recommendation feature, the object feature and the interest feature, the server may input the product recommendation feature, the object feature and the interest feature into a product recommendation model, splice the product recommendation feature, the object feature and the interest feature through an MLP (multi-layer perceptron neural network) layer of the product recommendation model to obtain a spliced feature, and predict an interest score of the candidate object on the target product based on the spliced feature.

S208, selecting a target object with the interest score reaching a scoring condition from the candidate objects.

Wherein the scoring condition may be a score ranking condition or a score threshold condition.

In one embodiment, when the score condition is a score ranking condition, after obtaining the interest score of each candidate object on the target product, the server ranks each candidate object according to the interest score to obtain a ranked result, and selects a target object reaching the score ranking condition from each candidate object based on the ranked result.

For example, the respective candidate objects are arranged in descending order according to the interest score, and the candidate objects ranked in the top 5 are determined as target objects, or the candidate objects ranked in the top 20% are determined as target objects.

In one embodiment, the scoring condition is a score threshold condition, and the server selects a target object with an interest score greater than the score threshold from the candidate objects after obtaining the interest score of each candidate object for the target product.

Specifically, after obtaining the interest scores of the candidate objects on the target products, the server determines whether the interest scores of the candidate objects are greater than a score threshold, if the interest scores of the candidate objects are greater than the score threshold, the server determines that the candidate objects are target objects, and if the interest scores of the candidate objects are not greater than the score threshold, the server determines that the candidate objects are non-target objects.

S210, pushing product recommendation information of the target product to the target object.

Specifically, after determining the target object, the server sends product recommendation information of the target product to a terminal corresponding to the target object, so that the terminal corresponding to the target object displays the target product based on the received product recommendation information.

In one embodiment, after determining the target object, the server may further obtain product display configuration information of a terminal corresponding to each target object, adjust product recommendation information of the target product according to the product display configuration information, obtain adjusted product recommendation information of the target product, and send the adjusted product recommendation information of the target product to the terminal corresponding to the target object, so that the terminal corresponding to the target object displays the target product based on the received adjusted product recommendation information.

For example, the terminal product display configuration information includes the size of the product image display window, and after the server obtains the size of the product image display window of the terminal, the server can adjust the size of the product image of the target product according to the size so as to adapt to the size of the product image display window of the terminal, and send the product image of the target product after the size adjustment to the terminal corresponding to the target object, so that the terminal corresponding to the target object displays the received product image of the target product.

In the above embodiment, the server obtains the product recommendation feature of the target product, the object feature of the candidate object, and the behavior feature sequence obtained by product interaction of the candidate object under each product interaction scene; the interest characteristics of the candidate object to the target product are determined based on the behavior characteristics and the product recommendation characteristics in the behavior characteristic sequence, so that the interest of the candidate object in each scene can be comprehensively considered to obtain the real preference of the candidate object, the interest score of the candidate object to the target product is determined according to the product recommendation characteristics, the object characteristics and the splicing characteristics among the interest characteristics, the target object with the interest score reaching the score condition is selected from the candidate object, and the accuracy of product recommendation is improved when the product recommendation information of the target product is pushed to the target object.

In one embodiment, after obtaining the behavior information set obtained by product interaction of the candidate object under each product interaction scene, the server may further sort the behavior information in the behavior information set according to the behavior time, and perform vectorization processing on the product recommendation information, the object information, and each behavior information in the sorted behavior information set.

The ranked behavior information set may also be referred to as a behavior information sequence, where the behavior information sequence refers to a sequence obtained by ranking behavior information according to behavior time.

It can be understood that the behavior habit or preference of the object is carried in the sequence of the behavior time of the behavior information, so that the behavior information sequence is obtained by sequencing the behavior information, and the prediction result can be more accurate when the interest score of the object to the target product is predicted based on the behavior information sequence.

Specifically, vectorization processing is carried out on each behavior information in the server product recommendation information, the object information and the ordered behavior information set to obtain a product recommendation information vector, an object information vector and a behavior information vector sequence in a high-dimensional space, and dimension reduction processing is carried out on the product recommendation information vector, the object information vector and the behavior information vector sequence in the high-dimensional space by adopting a preset dimension reduction algorithm to obtain product recommendation features, object features and behavior feature sequences.

In the above embodiment, the server ranks the behavior information in the behavior information set according to the behavior time, so as to obtain the behavior information sequence, where the sequence of the behavior time of the behavior information carries the behavior habit or preference of the candidate object, so that the behavior information sequence is obtained by ranking the behavior information, and when the interest score of the object to the target product is predicted based on the behavior information sequence, the prediction result can be more accurate.

In one embodiment, after the server acquires the behavior feature sequence, the server may further perform feature extraction on the behavior feature sequence based on the multi-head attention mechanism to obtain a corresponding attention behavior feature.

The number of attention behavior features corresponding to the same behavior feature is consistent with the number of the attention heads of the multi-head attention mechanism, and the number of the attention heads refers to the number of the attention heads of the multi-head attention mechanism. The multi-headed attention mechanism is used to model internal relationships between behavior feature sequences.

Specifically, after the server acquires the behavior feature sequence, determining attention matrixes corresponding to all attention heads of the multi-head attention mechanism based on the behavior feature sequence, and calculating attention behavior feature sequences corresponding to all attention heads according to the behavior feature sequence and the attention moment matrixes corresponding to all attention heads.

For example, the number of attention heads of the multi-head attention mechanism is N, and the attention behavior feature sequence corresponding to the h-th attention head determined based on the behavior feature sequence is:

wherein x is _b Head for behavior feature sequence _h For the attention behavior feature sequence corresponding to the h attention header,

and->

For the attention matrix corresponding to the h attention head, softmax is the activation function, +.>

Scaling factors.

In one embodiment, after obtaining the attention behavior feature sequences corresponding to the attention headers, the server may further determine an attention behavior feature sequence corresponding to the behavior feature sequence according to the attention behavior feature sequences corresponding to the attention headers.

Specifically, the server may determine the attention behavior feature sequence corresponding to the behavior feature sequence using the following formula:

Z＝MultiHead(x _b )＝concat(head ₁ ,…,head _h ,…,head _N )W ^o (6)

wherein Z is the attention lineFor characteristic sequences, head _h Attention behavior feature sequence corresponding to the h attention head, W ^o Is a parameter matrix which can be learned.

In one embodiment, the step of determining sub-interest features corresponding to each behavior feature by the server based on the behavior feature, the behavior time, the position code, and the product recommendation feature in the behavior feature sequence includes: aiming at each attention behavior feature sequence, determining sub-interest features corresponding to the current attention behavior feature sequence based on the current attention behavior feature sequence, the product recommendation feature of the target product, the behavior time and the position code until the sub-interest features corresponding to each attention behavior feature are obtained; and splicing sub-interest features of the attention behavior features corresponding to each behavior feature to obtain sub-interest features corresponding to each behavior feature.

It will be appreciated that the individual attention behavior features in the attention behavior feature sequence for each attention header correspond to behavior features in the behavior feature sequence, e.g., behavior feature sequence x _b ＝{e ₁ ,e ₂ ,…,e _i ,…,e _n The attention behavior feature sequence of the corresponding h attention head is head _h ＝{I _1,h ,I _2,h ,…,I _i,h ,…,I _n,h }, wherein the attention is directed to feature I _1,h The corresponding behavior is characterized as e ₁ Attention features I _2,h The corresponding behavior is characterized as e ₂ Attention features I _i,h The corresponding behavior is characterized as e _i Attention features I _n,h The corresponding behavior is characterized as e _n . It can be seen that when the behavior feature sequence x _b ＝{e ₁ ,e ₂ ,…,e _i ,…,e _n When N corresponding attention heads are provided, the behavior characteristic e _i N corresponding attention features are respectively I _i,1 、I _i,2 ……I _i,h ……I _i,N 。

Specifically, for any attention behavior feature in any attention behavior feature sequence, the server determines the time difference of the current time of the time interval of the behaviors according to the behavior time of the behavior feature corresponding to the attention behavior feature, discretizes the time difference to obtain a time coefficient, and determines the sub-interest feature corresponding to the attention behavior feature based on the time coefficient, the attention behavior feature, the position code and the product recommendation feature through the following sub-interest feature determining function. Wherein the expression of the sub-interest feature determination function is as follows:

Interest _i,h ＝α _i (I _i,h ,x _u ,p _i )×I _i,h (7)

Wherein, the Intrest _i,h For the ith attention behavior feature I in the h attention behavior feature sequence _i,h Corresponding sub-interest feature, alpha _i For the ith attention behavior feature I _i,h Corresponding behavior feature e _i Time coefficient, x _u Product recommendation features for target products, p _i For the ith attention behavior feature I _i,h Corresponding behavior feature e _i Is a position code of (a).

In one embodiment, after obtaining sub-interest features corresponding to each attention behavior feature in any attention behavior feature sequence, the server superimposes each sub-interest feature in the attention behavior feature sequence, so as to obtain a sub-interest feature corresponding to the attention behavior feature sequence, and specifically, the following formula may be adopted to determine the sub-interest feature corresponding to the attention behavior feature sequence:

wherein, the Intrest _h For the sub-Interest feature corresponding to the h attention behavior feature sequence, interest _i,h For the ith attention behavior feature I in the h attention behavior feature sequence _i,h The corresponding sub-interest features, n, are the number of behavioral features.

In an embodiment, after obtaining sub-interest features corresponding to each attention behavior feature in each attention behavior feature sequence, that is, after determining sub-interest features corresponding to each attention behavior feature corresponding to any one behavior feature, the server may further splice sub-interest features of each attention behavior feature corresponding to the behavior feature to obtain sub-interest features corresponding to each behavior feature. The following formula may be specifically adopted to determine sub-interest features corresponding to the behavior features:

Interest _i ＝(Interest _i,1 ,…,Interest _i,h ,…,Interest _i,N ) (9)

Wherein, the Intrest _i Ith behavior feature e _i Corresponding sub-Interest feature, interest _i,h Ith behavior feature e _i Sub-interest features of the corresponding h attention behavior feature.

In the above embodiment, the server performs feature extraction on the behavior feature sequence based on the multi-head attention mechanism to obtain the corresponding attention behavior feature, so as to realize the fusion of the behavior features, further determine sub-interest features corresponding to each behavior feature based on the attention behavior feature obtained after fusion, and determine the interest feature of the candidate object on the target product according to each sub-interest feature, namely, the method realizes the determination of the interest feature of the candidate object on the target product based on the correlation between the product in the fused behavior feature and the target product, and further can further improve the accuracy of product pushing when the product pushing is determined based on the interest feature.

In one embodiment, the server may further obtain context information formed by product interaction of the candidate object in each product interaction scene, and sequentially perform vectorization processing and dimension reduction processing on the context information to obtain the context feature.

The context information may be a content context, a time context, or a location context, for example, the content context is a context of introduction information of an object accessing an object, the time context is a life cycle of the object, and the location context is a model of a terminal used when accessing the object. Vectorization refers to the conversion of information into high-dimensional vectors for processing the information through an artificial intelligence model, and dimension reduction refers to the reduction of redundant information for information in the form of high-dimensional vectors to obtain low-dimensional vectors so as to improve the subsequent processing speed of the information.

It can be understood that the content, time and place often have a certain influence on the behavior of the candidate object, that is, the context information also carries the behavior habit or preference of the candidate object, so that the prediction result can be more accurate by acquiring the context information and extracting the context characteristics, and further predicting the interest score of the object in the target product based on the context characteristics.

Specifically, after obtaining context information formed by product interaction of candidate objects in each product interaction scene, the server performs vectorization processing on the context information to obtain context information vectors in a high-dimensional space, and then performs dimension reduction processing on the context information vectors in the high-dimensional space by adopting a preset dimension reduction algorithm to obtain context characteristics.

In one embodiment, the process of determining the interest score of the candidate object for the target product based on the splice features between the product recommendation feature, the object feature, and the interest feature by the server comprises the steps of: splicing the product recommendation features, the object features, the interest features and the context features to obtain splicing features; and determining the interest score of the candidate object to the target product based on the splicing characteristics.

Specifically, after obtaining the product recommendation feature, the object feature, the interest feature and the context feature, the server may input the product recommendation feature, the object feature, the interest feature and the context feature into a product recommendation model, splice the product recommendation feature, the object feature, the interest feature and the context feature through an MLP (multi-layer perceptron neural network) layer of the product recommendation model to obtain a spliced feature, and predict an interest score of the candidate object on the target product based on the spliced feature.

In the above embodiment, the server obtains the context characteristics formed by product interaction of the candidate objects in each product interaction scene, so that the prediction result can be more accurate when predicting the interest score of the object in the target product based on the context characteristics.

In one embodiment, the product recommendation feature, the object feature, the behavior feature sequence and the context feature are obtained by performing feature processing on a product recommendation model, and the recommendation method of the target product further includes a process of training the product recommendation model, as shown in fig. 3, where the process specifically includes the following steps:

s302, training product recommendation information corresponding to the sample product, training object information corresponding to the sample object and training behavior information sets obtained by the interaction of the sample object on the sample product under each product interaction scene are input into a product recommendation model to be trained, and training product recommendation characteristics, training object characteristics and training behavior characteristic sequences are obtained.

The product recommendation model comprises an embedding layer, wherein the embedding layer is used for carrying out dimension reduction processing on an input high-dimension vector.

Specifically, after obtaining training product recommendation information corresponding to a sample product, training object information corresponding to a sample object and training behavior information sets obtained by the sample object interacting with the sample product under each product interaction scene, a server carries out vectorization processing on the training product recommendation information, the training object information and the training behavior information sets to obtain training product recommendation vectors, training object vectors and training behavior vector sets in a high-dimensional space, then inputs the training product recommendation vectors, the training object vectors and the training behavior vector sets in the high-dimensional space into a product recommendation model to be trained, and carries out dimension reduction processing on the training product recommendation vectors, the training object vectors and the training behavior vector sets in the high-dimensional space by adopting a preset dimension reduction algorithm through an embedding (empedding) layer of the product recommendation model to be trained to obtain training product recommendation features, training object features and training behavior feature sequences.

S304, determining training interest characteristics of the sample object on the sample product based on the behavior characteristics in the training behavior characteristic sequence and the recommended characteristics of the training product.

In one embodiment, S304 specifically includes the steps of obtaining, by a server, a behavior time and a position code corresponding to a behavior feature in a training behavior feature sequence; determining sub-training interest features corresponding to the behavior features based on the behavior features, the behavior time, the position codes and the recommended features of the training products in the training behavior feature sequence; and determining training interest characteristics of the sample object on the sample product according to the sub-training interest characteristics.

The position coding refers to the coding sequence of the positions of the training behavior features in the training behavior feature sequence.

Specifically, after obtaining the behavior time and the position code corresponding to the behavior feature in the training behavior feature sequence, the server determines a time coefficient according to the behavior time, determines sub-training interest features corresponding to the behavior features according to the behavior features, the time coefficient, the position code and the training product recommendation feature, and determines training interest features of the sample object on the sample product according to the sub-training interest features corresponding to the behavior features.

S306, according to the recommended features of the training products, the features of the training objects and the spliced features between the training interest features, the interest scores of the sample objects on the sample products are determined.

Specifically, after obtaining the recommended feature of the training product, the feature of the training object and the feature of the training interest, the server may input the recommended feature of the training product, the feature of the training object and the feature of the training interest into an MLP (multi-layer perceptron neural network) layer of the product recommendation model, splice the recommended feature of the training product, the feature of the training object and the feature of the training interest through the MLP (multi-layer perceptron neural network) layer to obtain a spliced feature, and determine an interest score of the sample object on the sample product based on the spliced feature.

S308, determining a first loss value between the interest score and the interest tag.

The interest labels refer to real labels of samples, and represent sample products with which sample objects actually interact.

Specifically, after obtaining the interest score, the server may further obtain an interest tag corresponding to the interest score, and determine a first loss value based on the interest score and the interest tag through a first loss function, where an expression of the first loss function is as follows:

wherein L is _target For the first loss value, M and T are both the total number of samples, i.e., m=t, y _i For interest tags of the sample, i.e. desired output, f (x _i ) Is the actual output of the model, i.e., the score of interest.

S310, adjusting parameters of a product recommendation model to be trained based on the first loss value.

Specifically, after obtaining the first loss value, the server adjusts parameters of the product recommendation model to be trained based on the first loss value to obtain an adjusted product recommendation model, and then re-executes step S302 until the model converges, and determines the adjusted product recommendation model obtained at this time as a trained product recommendation model.

In the above embodiment, the server trains the product recommendation model based on training product recommendation information corresponding to the sample product, training object information corresponding to the sample object, and training behavior information sets obtained by the sample object interacting with the sample product in each product interaction scene, so that the trained product recommendation model has the capability of extracting interests of candidate objects in each scene to obtain real favorites of the candidate objects, and further, when product recommendation is performed based on the product recommendation model, accuracy of product recommendation is improved.

In one embodiment, the process of training the product recommendation model by the server further comprises the steps of: based on the behavior characteristics in the training behavior characteristic sequence, determining the predicted behavior characteristics corresponding to the sample object at the next moment; a second loss value is determined based on the predicted behavioral characteristics and the behavioral characteristics signature.

Wherein the second loss value may also be referred to as an auxiliary loss value. The behavior feature tag refers to a real behavior tag of the sample object at the next moment, and represents the real behavior of the sample object at the next moment.

Specifically, after obtaining behavior features in the training behavior sequence, the server predicts predicted behavior features corresponding to the sample object at the next moment based on at least one behavior feature in the obtained behavior features, obtains a behavior feature tag, and determines a first loss value based on the predicted behavior features and the behavior feature tag through a second loss function, wherein the expression of the second loss function is as follows:

wherein sigma (-) represents a sigmoid activation function,<.>represent the inner product, L _aux For the second loss value, M is the total number of samples,

for the behavior characteristic of the ith sample at time t, < >>

Predicted behavior feature of the ith sample at time t+1,/for example>

And (3) marking the behavior characteristic of the ith sample at the time t+1. It will be appreciated that time t represents the current time and time t+1 represents the time next to the current time.

In one embodiment, after obtaining the first loss value and the second loss value, the server may further perform parameter adjustment on the product recommendation model to be trained based on the first loss value and the second loss value.

Specifically, the server determines a model loss value of the product recommendation model to be trained based on the first loss value and the second loss value, adjusts parameters of the product recommendation model to be trained according to the training loss value to obtain an adjusted product recommendation model, and then re-executes step S302 until the model converges, and determines the adjusted product recommendation model obtained at this time as a trained product recommendation model.

In the above embodiment, the server determines the predicted behavior feature corresponding to the sample object at the next time based on the behavior feature in the training behavior feature sequence, determines the second loss value based on the predicted behavior feature and the behavior feature label, and further adjusts the parameters of the product recommendation model to be trained based on the first loss value and the second loss value, so that the trained product recommendation model can better extract the interests of the candidate object in each scene to obtain the real preference of the candidate object, and further improves the accuracy of product recommendation when the product recommendation is performed based on the product recommendation model.

In one embodiment, the process of the server for parameter adjustment of the product recommendation model to be trained based on the first loss value and the second loss value comprises the steps of: obtaining a loss coefficient corresponding to the second loss value; determining a model loss value according to the loss coefficient, the first loss value and the second loss value; and carrying out parameter adjustment on the product recommended model to be trained based on the model loss value, and stopping training until the obtained model loss value reaches a preset condition.

Wherein the loss factor of the second loss value is used to adjust the weight of the second loss value in the model loss value, i.e. the bar or the influence of the second loss value on the model loss value.

Specifically, after obtaining the loss coefficient corresponding to the second loss value, the server determines a model loss value based on the loss coefficient, the first loss value and the second loss value through a third loss function, adjusts parameters of the product recommendation model to be trained based on the model loss value to obtain an adjusted product recommendation model, and then re-executes step S302 until the obtained model loss value reaches a preset condition, and determines the adjusted product recommendation model obtained at this time as a trained product recommendation model.

Wherein the preset condition may be that the model converges or that the model loss value reaches a target loss value. The expression of the third loss function is as follows:

L _total ＝L _target +λL _aux (12)

wherein L is _total For model loss value, L _target As a result of the first loss value,L _aux for the second loss value, λ is the loss coefficient of the second loss value.

In the above embodiment, the server determines the model loss value according to the loss coefficient corresponding to the second loss value, and the first loss value and the second loss value, and adjusts parameters of the product recommendation model to be trained based on the model loss value until the obtained model loss value reaches a preset condition, so that the trained product recommendation model can better extract interests of the candidate object in each scene to obtain real favorites of the candidate object, and further improve accuracy of product recommendation when product recommendation is performed based on the product recommendation model.

In one embodiment, the process of determining the predicted behavior feature corresponding to the sample object at the next time by the server based on the behavior feature in the training behavior feature sequence specifically includes the following steps: feature extraction is carried out on the training behavior feature sequence based on a multi-head attention mechanism, so that corresponding attention behavior features are obtained; and determining the predicted behavior characteristics corresponding to the sample object at the next moment based on the attention behavior characteristics. The number of attention behavior features corresponding to each behavior feature is consistent with the number of heads of the multi-head attention mechanism.

Specifically, the server may perform feature extraction on the training behavior feature sequence based on the multi-head attention mechanism to obtain attention behavior feature sequences corresponding to all attention heads, determine attention behavior feature sequences corresponding to the behavior feature sequences according to the attention behavior feature sequences corresponding to all attention heads, and determine predicted behavior features corresponding to sample objects at the next moment according to all attention behavior features corresponding to the same sample product in the obtained attention behavior feature sequences.

In one embodiment, the server determines a predicted behavior feature corresponding to the sample object at a next time based on the attention behavior feature, and may further obtain a behavior feature tag, and determine a first loss value based on the predicted behavior feature and the behavior feature tag through a second loss function, where an expression of the second loss function is as follows:

for the attention behavior characteristic of the ith sample at time t,/for example>

The predicted behavior characteristics of the ith sample at time t +1,

In the above embodiment, the server performs feature extraction on the training behavior feature sequence based on the multi-head attention mechanism to obtain the corresponding attention behavior feature, determines the predicted behavior feature corresponding to the sample object at the next moment based on the attention behavior feature, determines the second loss value based on the predicted behavior feature and the behavior feature label, and further performs parameter adjustment on the product recommendation model to be trained based on the first loss value and the second loss value, so that the trained product recommendation model has better capability of fusing the behavior feature, further can better extract interests of the candidate object in each scene to obtain real favors of the candidate object, and further improves accuracy of product recommendation when product recommendation is performed based on the product recommendation model.

In one embodiment, when the server trains the product recommendation model to be trained, training context information formed by interaction of the sample object with the sample product in each product interaction scene can be obtained; and inputting the training context information into a product recommendation model to be trained to obtain training context characteristics.

The context information may be a content context, a time context, or a location context, for example, the content context is a context of introduction information of an object accessing an object, the time context is a life cycle of the object, and the location context is a model of a terminal used when accessing the object.

Specifically, after the server obtains the context information, the context information may be input into a product recommendation model to be trained, and the vectorization processing and the dimension reduction processing are sequentially performed on the context information through the product recommendation model to obtain the context feature.

The vectorization means that the information is converted into a high-dimensional vector, the information is processed through an artificial intelligent model, and the dimension reduction process means that redundant information is reduced for the information in the form of the high-dimensional vector to obtain a low-dimensional vector so as to improve the subsequent processing speed of the information.

In one embodiment, the process of determining the interest score of the sample object for the sample product by the server based on the recommended feature of the training product, the stitching feature between the training object feature and the training interest feature further comprises the steps of: and determining the interest scores of the sample objects on the sample products according to the recommended features of the training products, the features of the training objects, the training interest features and the splicing features among the training context features.

Specifically, after obtaining the recommended feature of the training product, the feature of the training object, the feature of the training interest and the feature of the training context, the server may input the recommended feature of the training product, the feature of the training object, the feature of the training interest and the feature of the training context into a product recommendation model, splice the recommended feature of the training product, the feature of the training object, the feature of the training interest and the feature of the training context through an MLP (multi-layer perceptron neural network) layer of the product recommendation model to obtain a spliced feature, and predict the interest score of the sample object on the sample product based on the spliced feature.

In the above embodiment, the server acquires the training context information formed by the interaction of the sample object with the sample product in each product interaction scene, and further trains the product recommendation model based on the context information, so that the trained product recommendation model can make the prediction result more accurate when predicting the interest score of the object to the target product based on the context characteristics.

In one embodiment, as shown in fig. 4, there is further provided a recommendation method for a target product, which is illustrated by using the method applied to the server in fig. 1 as an example, and includes the following steps:

s402, obtaining product recommendation information of a target product, object information of a candidate object, a behavior information set obtained by product interaction of the candidate object under each product interaction scene, and context information formed by product interaction of the candidate object under each product interaction scene.

S404, sorting the behavior information in the behavior information set according to the behavior time.

S406, vectorizing the product recommendation information, the object information, the behavior information in the ordered behavior information set and the context information.

S408, performing dimension reduction processing on the vectorized result to obtain product recommended features, object features, behavior feature sequences and context features.

S410, acquiring the behavior time and the position code corresponding to the behavior feature in the behavior feature sequence.

And S412, extracting the characteristics of the behavior characteristic sequence based on a multi-head attention mechanism to obtain corresponding attention behavior characteristics.

The number of attention behavior features corresponding to each behavior feature is consistent with the number of heads of the multi-head attention mechanism.

S414, aiming at each attention behavior feature, determining a sub-interest feature corresponding to the current attention behavior feature based on the current attention behavior feature, the product recommendation feature of the target product, the behavior time and the position code until the sub-interest feature corresponding to each attention behavior feature is obtained.

S416, sub-interest features of the attention behavior features corresponding to each behavior feature are spliced, and sub-interest features corresponding to each behavior feature are obtained.

S418, according to each sub-interest feature, determining the interest feature of the candidate object to the target product.

S420, determining interest scores of candidate objects on target products according to the product recommendation characteristics, the object characteristics and the splicing characteristics among the interest characteristics.

S422, selecting a target object with the interest score reaching a scoring condition from the candidate objects.

S424, pushing product recommendation information of the target product to the target object.

The application scene is applied to the recommendation method of the target product, and the recommendation method of the target product is achieved through a trained product recommendation model. Specifically, the application of the recommendation method of the target product in the application scene is as follows:

Referring to the flowchart shown in fig. 5, a server acquires training samples for training a product recommendation model in advance, the training samples comprise training product recommendation information corresponding to sample products, training object information corresponding to sample objects and training behavior information sets obtained by the sample objects in interaction with the sample products under each product interaction scene, training data are generated according to the training samples, and a structure of the product recommendation model is constructed, and the product recommendation model comprises an Embedding layer (Embedding layer), a behavior extraction layer, a multi-sequence fusion layer and an MLP (multi-layer perceptron neural network) layer as shown in fig. 6; and training the product recommendation model by adopting the generated training data, predicting the interest score of the candidate object to the target product by the product recommendation model, and pushing the target advertisement to the target object with the interest score reaching the score condition, namely writing the target object into the target product for orientation.

It should be understood that, although the steps in the flowcharts of fig. 2-5 are shown in order as indicated by the arrows, these steps are not necessarily performed in order as indicated by the arrows. The steps are not strictly limited to the order of execution unless explicitly recited herein, and the steps may be executed in other orders. Moreover, at least some of the steps in fig. 2-5 may include multiple steps or stages that are not necessarily performed at the same time, but may be performed at different times, nor does the order in which the steps or stages are performed necessarily performed in sequence, but may be performed alternately or alternately with at least a portion of the steps or stages in other steps or other steps.

In one embodiment, as shown in fig. 7, a recommendation device for a target product is provided, where the recommendation device may use a software module or a hardware module, or a combination of the two, and the recommendation device is a part of a computer device, and specifically includes: a feature acquisition module 702, an interest feature determination module 704, an interest score determination module 706, an object selection module 708, and a product recommendation module 710, wherein:

the feature acquisition module 702 is configured to acquire product recommendation features of a target product, object features of a candidate object, and a behavior feature sequence obtained by product interaction of the candidate object in each product interaction scene.

The interest feature determining module 704 is configured to determine an interest feature of the candidate object for the target product based on the behavior feature and the product recommendation feature in the behavior feature sequence.

The interest score determining module 706 is configured to determine an interest score of the candidate object for the target product according to the product recommendation feature, the object feature, and the stitching feature between the interest features.

An object selection module 708 is configured to select a target object whose interest score reaches a score condition from the candidate objects.

The product recommendation module 710 is configured to push product recommendation information of a target product to a target object.

In the above embodiment, the product recommendation feature of the target product, the object feature of the candidate object, and the behavior feature sequence obtained by product interaction of the candidate object under each product interaction scene are obtained; the interest characteristics of the candidate object to the target product are determined based on the behavior characteristics and the product recommendation characteristics in the behavior characteristic sequence, so that the interest of the candidate object in each scene can be comprehensively considered to obtain the real preference of the candidate object, the interest score of the candidate object to the target product is determined according to the product recommendation characteristics, the object characteristics and the splicing characteristics among the interest characteristics, the target object with the interest score reaching the score condition is selected from the candidate object, and the accuracy of product recommendation is improved when the product recommendation information of the target product is pushed to the target object.

In one embodiment, the feature acquisition module 702 is further configured to: acquiring product recommendation information of a target product, object information of a candidate object and a behavior information set obtained by product interaction of the candidate object under each product interaction scene; vectorizing each piece of behavior information in the product recommendation information, the object information and the behavior information set; and performing dimension reduction treatment on the vectorized result to obtain product recommended features, object features and behavior feature sequences.

In one embodiment, as shown in fig. 8, the apparatus further comprises: a behavior information ordering module 712, configured to order behavior information in the behavior information set according to behavior time; the feature acquisition module 702 is further configured to: and carrying out vectorization processing on each piece of behavior information in the product recommendation information, the object information and the ordered behavior information set.

In one embodiment, the interest feature determination module 704 is further configured to: acquiring behavior time and position codes corresponding to behavior features in a behavior feature sequence; determining sub-interest features corresponding to the behavior features based on the behavior features, the behavior time, the position codes and the product recommendation features in the behavior feature sequence; and determining the interest characteristics of the candidate object on the target product according to the sub-interest characteristics.

In one embodiment, the interest feature determination module 704 is further configured to: extracting the characteristics of the behavior characteristic sequence based on a multi-head attention mechanism to obtain corresponding attention behavior characteristics; the number of attention behavior features corresponding to each behavior feature is consistent with the number of heads of the multi-head attention mechanism; aiming at each attention behavior feature, determining a sub-interest feature corresponding to the current attention behavior feature based on the current attention behavior feature, the product recommendation feature of the target product, the behavior time and the position code until the sub-interest feature corresponding to each attention behavior feature is obtained; and splicing sub-interest features of the attention behavior features corresponding to each behavior feature to obtain sub-interest features corresponding to each behavior feature.

In one embodiment, the feature acquisition module 702 is further configured to: acquiring context information formed by product interaction of candidate objects in each product interaction scene; carrying out vectorization processing and dimension reduction processing on the context information in sequence to obtain context characteristics; the interest score determination module 706 is further configured to: splicing the product recommendation features, the object features, the interest features and the context features to obtain splicing features; and determining the interest score of the candidate object to the target product based on the splicing characteristics.

In one embodiment, the product recommendation feature, the object feature, the behavioral feature sequence, and the contextual feature are obtained by feature processing through a product recommendation model; as shown in fig. 8, the apparatus further includes:

the training feature obtaining module 714 is configured to input training product recommendation information corresponding to a sample product, training object information corresponding to a sample object, and a training behavior information set obtained by the sample object interacting with the sample product under each product interaction scene, to a product recommendation model to be trained, to obtain a training product recommendation feature, a training object feature, and a training behavior feature sequence.

The training interest feature obtaining module 716 is configured to determine training interest features of the sample object on the sample product based on the behavior features in the training behavior feature sequence and the training product recommendation features.

The training interest score determining module 718 is configured to determine an interest score of the sample object for the sample product according to the recommended feature of the training product, the feature of the training object, and the stitching feature between the training interest features.

A loss value determination module 720 for determining a first loss value between the interest score and the interest tag.

The model parameter adjustment module 722 is configured to adjust parameters of a product recommendation model to be trained based on the first loss value.

In one embodiment, the loss value determination module 720 is further configured to: based on the behavior characteristics in the training behavior characteristic sequence, determining the predicted behavior characteristics corresponding to the sample object at the next moment; determining a second loss value based on the predicted behavioral characteristics and the behavioral characteristics signature; the model parameter adjustment module 722 is further configured to: and carrying out parameter adjustment on the product recommendation model to be trained based on the first loss value and the second loss value.

In one embodiment, the model parameter adjustment module 722 is further configured to: obtaining a loss coefficient corresponding to the second loss value; determining a model loss value according to the loss coefficient, the first loss value and the second loss value; and carrying out parameter adjustment on the product recommended model to be trained based on the model loss value, and stopping training until the obtained model loss value reaches a preset condition.

In one embodiment, training feature acquisition module 714 is further to: acquiring training context information formed by interaction of a sample object with a sample product in each product interaction scene; inputting training context information into a product recommendation model to be trained to obtain training context characteristics;

training interest score determination module 718 is further configured to: and determining the interest scores of the sample objects on the sample products according to the recommended features of the training products, the features of the training objects, the training interest features and the splicing features among the training context features.

The specific limitation of the recommendation device for the target product may be referred to the limitation of the recommendation method for the target product hereinabove, and will not be described herein. The above-mentioned respective modules in the recommendation device for the target product may be implemented in whole or in part by software, hardware, and a combination thereof. The above modules may be embedded in hardware or may be independent of a processor in the computer device, or may be stored in software in a memory in the computer device, so that the processor may call and execute operations corresponding to the above modules.

In one embodiment, a computer device is provided, which may be a server, and the internal structure of which may be as shown in fig. 9. 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 includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database of the computer device is used to store product information and object data. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program, when executed by the processor, implements a method of recommending a target product.

In one embodiment, a computer device is provided, which may be a terminal, and an internal structure diagram thereof may be as shown in fig. 10. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device 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 includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The communication interface of the computer device is used for carrying out wired or wireless communication with an external terminal, and the wireless mode can be realized through WIFI, an operator network, NFC (near field communication) or other technologies. The computer program, when executed by the processor, implements a method of recommending a target product. The display screen of the computer equipment can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment can be a touch layer covered on the display screen, can also be keys, a track ball or a touch pad arranged on the shell of the computer equipment, and can also be an external keyboard, a touch pad or a mouse and the like.

It will be appreciated by those skilled in the art that the structures shown in fig. 9 or 10 are merely block diagrams of portions of structures related to the present application and do not constitute a limitation of the computer device on which the present application may be applied, and that a particular computer device may include more or fewer components than shown, or may combine certain components, or have a different arrangement of components.

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

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

In one embodiment, a computer program product or computer program is provided that includes computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the steps in the above-described method embodiments.

Those skilled in the art will appreciate that implementing all or part of the above described methods may be accomplished by way of a computer program stored on a non-transitory computer readable storage medium, which when executed, may comprise the steps of the embodiments of the methods described above. Any reference to memory, storage, database, or other medium used in embodiments provided herein may include at least one of non-volatile and volatile memory. The nonvolatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, or the like. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory. By way of illustration, and not limitation, RAM can be in the form of a variety of forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), and the like.

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

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

Claims

1. A method for recommending a target product, the method comprising:

2. The method according to claim 1, wherein the obtaining product recommendation features of the target product, object features of the candidate object, and a behavior feature sequence obtained by product interaction of the candidate object in each product interaction scene includes:

acquiring product recommendation information of a target product, object information of a candidate object and a behavior information set obtained by product interaction of the candidate object under each product interaction scene;

vectorizing each piece of behavior information in the product recommendation information, the object information and the behavior information set;

and performing dimension reduction on the vectorized result to obtain the product recommended feature, the object feature and the behavior feature sequence.

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

sorting the behavior information in the behavior information set according to the behavior time;

the vectorizing the product recommendation information, the object information and the behavior information in the behavior information set includes:

4. The method of claim 1, wherein the determining the candidate object's interest feature in the target product based on the behavioral features in the behavioral feature sequence and the product recommendation feature comprises:

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

the determining sub-interest features corresponding to the behavior features based on the behavior features in the behavior feature sequence, the behavior time, the position codes and the product recommendation features includes:

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

the determining the interest score of the candidate object to the target product according to the product recommendation feature, the object feature and the splicing feature between the interest features comprises the following steps:

7. The method according to any one of claims 1 to 6, wherein the product recommendation feature, the object feature, the behavioral feature sequence, and the contextual feature are obtained by feature processing by a product recommendation model; the method further comprises the steps of:

training product recommendation information corresponding to a sample product, training object information corresponding to a sample object and a training behavior information set obtained by the interaction of the sample object on the sample product under each product interaction scene are input into the product recommendation model to be trained, so that training product recommendation characteristics, training object characteristics and training behavior characteristic sequences are obtained;

determining training interest features of the sample object on the sample product based on the behavior features in the training behavior feature sequence and the training product recommendation features;

determining the interest score of the sample object on the sample product according to the recommended feature of the training product, the characteristic of the training object and the splicing characteristic between the training interest characteristics;

determining a first loss value between the interest score and an interest tag;

And adjusting parameters of the product recommendation model to be trained based on the first loss value.

8. The method of claim 7, wherein the method further comprises:

the adjusting parameters of the product recommendation model to be trained based on the first loss value comprises:

9. The method of claim 8, wherein the parameter adjustment of the product recommendation model to be trained based on the first loss value and the second loss value comprises:

obtaining a loss coefficient corresponding to the second loss value;

10. The method according to any one of claims 7 to 9, characterized in that the method further comprises:

the determining the interest score of the sample object to the sample product according to the recommended feature of the training product, the feature of the training object and the splicing feature between the training interest features comprises:

11. A recommendation device for a target product, the device comprising:

12. The apparatus of claim 11, wherein the feature acquisition module is further configured to:

13. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any one of claims 1 to 10 when the computer program is executed.

14. A computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the steps of the method of any one of claims 1 to 10.

15. A computer program product comprising a computer program, characterized in that the computer program, when executed by a controller, implements the steps of the method of any one of claims 1 to 10.