CN115510313A

CN115510313A - Information recommendation method and device, storage medium and computer equipment

Info

Publication number: CN115510313A
Application number: CN202110698370.7A
Authority: CN
Inventors: 王良栋; 丘志杰; 刘书凯; 张博; 饶君
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2021-06-23
Filing date: 2021-06-23
Publication date: 2022-12-23

Abstract

The embodiment of the application discloses an information recommendation method, an information recommendation device, a storage medium and computer equipment, wherein the method comprises the following steps: extracting long-term interest features of the user according to the user portrait information; extracting short-term interest characteristics of the user according to click sequence information in the historical behavior data of the user; calculating an item attribute preference vector of the user according to item attribute information in the historical behavior data of the user; fusing the long-term interest characteristics, the short-term interest characteristics and the item attribute preference information of the user to obtain a user vector; optimizing a plurality of targets in the target model according to the user vector to obtain an optimized target model; the item information in the item information base is processed through the optimized target model to generate recall information corresponding to the target user, and recommendation is performed based on the recall information, so that the model can better reflect the interest preference of the user, and the recommendation accuracy of the model is improved.

Description

Information recommendation method and device, storage medium and computer equipment

Technical Field

The application relates to the technical field of artificial intelligence, in particular to an information recommendation method, an information recommendation device, a storage medium and computer equipment.

Background

The recommendation system is one of important applications in the field of artificial intelligence, and can help users find item information which may be interested in the users in an information overload environment and push the item information to the users who are interested in the items.

In a complex recommendation system, there are usually many targets to be optimized, which results in a great deviation between the recommended content of the recommendation system and the user interest preference.

Disclosure of Invention

The embodiment of the application provides an information recommendation method, an information recommendation device, a storage medium and computer equipment, which can aggregate user portrait information, click sequence information and item attribute information into a user vector representing interest preference of a user, and simultaneously optimize a plurality of targets in a target model based on the user vector, so that the model can better reflect the interest preference of the user, and recommendation accuracy of the model is improved.

In a first aspect, an information recommendation method is provided, where the method includes:

extracting long-term interest features of the user according to the user portrait information;

extracting short-term interest characteristics of the user according to click sequence information in the historical behavior data of the user;

calculating an item attribute preference vector of the user according to item attribute information in the historical behavior data of the user;

fusing the long-term interest characteristics, the short-term interest characteristics and the item attribute preference information of the user to obtain a user vector;

optimizing a plurality of targets in a target model according to the user vector to obtain an optimized target model;

and processing the item information in the item information base through the optimized target model to generate recall information corresponding to the target user, and recommending based on the recall information.

In a second aspect, an information recommendation apparatus is provided, the apparatus comprising:

the first extraction unit is used for extracting long-term interest characteristics of the user according to the user portrait information;

the second extraction unit is used for extracting the short-term interest characteristics of the user according to the click sequence information in the historical behavior data of the user;

the computing unit is used for computing item attribute preference vectors of the users according to item attribute information in the historical behavior data of the users;

the fusion unit is used for fusing the long-term interest characteristics, the short-term interest characteristics and the item attribute preference information of the user to obtain a user vector;

the optimization unit is used for optimizing a plurality of targets in a target model according to the user vector to obtain an optimized target model;

and the recommending unit is used for processing the item information in the item information base through the optimized target model to generate recall information corresponding to the target user and recommending the recall information.

In a third aspect, a computer-readable storage medium is provided, which stores a computer program, where the computer program is suitable for being loaded by a processor to execute the steps in the information recommendation method according to any one of the above embodiments.

In a fourth aspect, a computer device is provided, where the computer device includes a processor and a memory, and the memory stores a computer program, and the processor is configured to execute the steps in the information recommendation method according to any one of the above embodiments by calling the computer program stored in the memory.

The embodiment of the application provides an information recommendation method, an information recommendation device, a storage medium and computer equipment, wherein long-term interest characteristics of a user are extracted according to user portrait information; extracting short-term interest characteristics of the user according to click sequence information in the historical behavior data of the user; calculating an item attribute preference vector of the user according to item attribute information in the historical behavior data of the user; fusing the long-term interest characteristics, the short-term interest characteristics and the item attribute preference information of the user to obtain a user vector; optimizing a plurality of targets in the target model according to the user vector to obtain an optimized target model; and processing the item information in the item information base through the optimized target model to generate recall information corresponding to the target user, and recommending based on the recall information. According to the method and the device, the user portrait information, the click sequence information and the project attribute information are aggregated into the user vector representing the interest preference of the user, and the multiple targets in the target model are optimized simultaneously based on the user vector, so that the interest preference of the user can be reflected by the model, and the recommendation accuracy of the model is improved.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.

Fig. 1 is a schematic view of an application scenario of a recommendation system according to an embodiment of the present application.

Fig. 2 is a schematic flow chart of an information recommendation method provided in an embodiment of the present application.

Fig. 3 is a first framework diagram of an object model according to an embodiment of the present disclosure.

Fig. 4 is a second framework diagram of the target model provided in the embodiment of the present application.

Fig. 5 is a schematic structural diagram of an information recommendation device according to an embodiment of the present application.

Fig. 6 is a schematic structural diagram of a computer device according to an embodiment of the present application.

Detailed Description

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.

The embodiment of the application provides an information recommendation method, an information recommendation device, computer equipment and a storage medium. Specifically, the information recommendation method according to the embodiment of the present application may be executed by a computer device, where the computer device may be a terminal or a server. The terminal can be a smart phone, a tablet Computer, a notebook Computer, a smart television, a smart speaker, a wearable smart device, a Personal Computer (PC), and the like, and the terminal can further include a client, which can be a video client, a browser client, an instant messaging client, and the like. The server may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing basic cloud computing services such as cloud service, a cloud database, cloud computing, a cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, content Delivery Network (CDN), big data, an artificial intelligence platform, and the like.

Before further detailed description of the embodiments of the present application, terms and expressions referred to in the embodiments of the present application will be described, and the terms and expressions referred to in the embodiments of the present application will be used for the following explanation.

Artificial Intelligence (AI) is a theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human Intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a manner similar to human intelligence. Artificial intelligence is the research of the design principle and the realization method of various intelligent machines, so that the machines have the functions of perception, reasoning and decision making.

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

Natural Language Processing (NLP): an important direction in the fields of computer science and artificial intelligence can realize effective communication between people and computers by using natural language. Natural language processing is a science integrating linguistics, computer science and mathematics. Therefore, the field relates to natural language, namely the language which people use everyday, so that the field is closely related to linguistics. Natural language processing techniques typically include text processing, semantic understanding, machine translation, robotic question and answer, knowledge mapping, and the like.

Item (Item): is an article, a video item, or an audio item, etc.

Collaborative Filtration (CF): the method is characterized in that the information which is interesting to the user is recommended by utilizing the preference of a group which is interesting and has common experience, and the individual gives a considerable response (such as scoring) to the information through a cooperative mechanism and records the response so as to achieve the purpose of filtering and further help others to filter the information.

Multi-task learning: by training several tasks simultaneously, multiple tasks are made to influence each other, and share a structure, and parameters in the structure are influenced by all the tasks during optimization. Thus, when all tasks converge, the structure is equivalent to fusing all tasks, so that the generalization capability of multi-task learning is generally better than that of a single task.

ItemCF: similarity of items is calculated by mining co-occurrence information of the items, and then recommendation and filtering are performed by using the similarity of the items.

Item2Vec: each Item is assigned a dense vector, which, in contrast to the one-hot data representation approach, can preserve semantic dimension information from Item to Item.

Recurrent Neural Network (RNN): the multilayer feedback neural network is an artificial neural network with nodes directionally connected into a ring. The internal state of the RNN network may exhibit dynamic timing behavior. Unlike feed-forward neural networks, the RNN can use its internal memory to process input sequences of arbitrary timing, which allows the RNN to more easily handle, for example, unsegmented handwriting recognition, speech recognition, etc.

UserCF: and (3) mining User (User) similarity information and recommending articles liked by similar users.

LSTM is a long-short term memory network, is a time-recursive neural network, and is suitable for processing and predicting important events with relatively long intervals and delays in a time sequence.

Gated recycling Unit (Gated Recurrent Unit, GRU): by introducing a Reset Gate (Reset Gate) and an Update Gate (Update Gate), the propagation of history information can be controlled.

Expert Network (Expert Network): the network structure designed for different tasks in multi-task learning is usually composed of multiple layers of fully connected networks and is used for extracting features of different tasks.

And a Gate (Gate) for controlling probability distribution of input weights of different expert networks, and combining the different expert networks by different weights and then inputting the different expert networks into different objective optimization towers.

Cosine similarity: cosine similarity, also known as cosine similarity, is an evaluation of similarity between two vectors by calculating cosine values of an angle between the two vectors. Cosine similarity maps vectors into a vector space, such as the most common two-dimensional space, according to coordinate values.

The target user: users currently using recommendation systems.

User portrait: the system is also called as a user role, and is an effective tool for sketching user appeal and design direction. User images are widely used in various fields, and in the course of actual operations, attributes and behaviors of users are often combined with expectations by words appearing shallowest and living closely to each other to serve as virtual representations of actual users.

Recall (Recall): and retrieving related documents from the document library, for example, roughly selecting a batch of commodities to be recommended for the user.

In a complex recommendation system, there are usually many targets to be optimized, such as click rate, share rate, playing completion degree, and other optimization targets. In order to optimize the click rate of the User, the ItemCF and the UserCF in the collaborative filtering CF calculate the similarity between the users and the similarity between the items through the co-occurrence between the User and the Item, and then personalized recommendation is carried out on the User. The co-occurrence generally refers to a phenomenon that information described by feature items of the same or different types in a document co-occur, for example, the feature items include external and internal features of the document, such as title, author, keyword, organization and the like, in a recommendation system, there is an interactive behavior between a user and an item, and each interactive behavior is recorded as one co-occurrence; there are various categories of interactive behaviors, such as clicking, sharing, playing, etc. Another method for calculating the similarity of the items is also Item2vec, and personalized recommendation is carried out by training the semantic relevance among the items. And some methods (RNN, GRU and the like) predict the click of the user at the next moment by sequence modeling, so that the click rate of the user is improved. The method can be used for improving the sharing rate, the playing completion degree and other targets of the user.

For example, in the recommendation system of application a, the recommendation system focuses differently on each target, and usually optimizes different targets by training a plurality of different sub-networks, so that the training tasks can be trained independently; after all the sub-networks are trained, different pre-estimated values can be predicted, then the different pre-estimated values are subjected to weighted combination to calculate the interest score of the User on the Item, and finally the candidate Item of K (TopK) before ranking is obtained through sorting. For example, in the recommendation system of application a, the click rate and the sharing rate of the user are not an order of magnitude, and the user usually has a higher click rate than the sharing rate of the user. Therefore, under a certain target recommendation scene, the problem of data sparseness exists, that is, the number of samples for training the click rate is far greater than the number of samples for training the share rate, the problem of data sparseness enables the final interest score of the user to be biased to the click rate model, and actually, the share rate of the user experiences the interest preference of the user.

The optimization independent training of the multiple targets does not conform to the actual scene, the targets to be optimized are usually related, for example, the click rate of the user and the sharing rate of the user are usually positively related, the sharing rate of the user is usually positively related to the playing completion degree of the video, and if the tasks of the targets to be optimized are trained separately, the relevance of the indexes is cut, so that the problem that the targets do not conform to the actual scene exists. In addition, separate training may cause another problem, the problem of sample selection bias. In a target recommendation scenario, the user's interest scores are generally calculated for all candidate items, where model training is performed on a small portion of a training set, and the samples of the training share rate are more in a subset of the samples of the training click, and model prediction is performed in the entire sample space, thereby generating a sample selection bias problem. That is, the distribution of the training set and the test set is different, which violates the assumption that the samples need to be independently distributed in machine learning.

Therefore, the embodiment of the application provides an information recommendation method, which is characterized in that a basic target model is designed, user portrait information, click sequence information and item attribute information are aggregated into a user vector, then a plurality of targets in the target model are optimized, and information recommendation is performed on a target user based on the optimized target model. Specifically, user portrait information is modeled as long-term interest characteristics of a user; modeling click sequence information as short-term interest characteristics of a user; and simultaneously, item attribute preference information of the user for the item is modeled by utilizing the item attribute information. The present application becomes a Multi-target simultaneous optimization compared to the way multiple targets are trained separately, where a Multi-gate mix-of-Experts (MMoE) module is employed to combine the impact of different expert networks on different target towers. For example, different target towers may include a share target and a play target.

The MMoE module extracts features of different directions through a plurality of expert networks, the expert networks generally consist of DNN structures, embedded structures can be shared at the bottoms of the expert networks, and each expert network can bring different influences on different tasks. And on the upper part of the expert network, towers with different optimization targets are used for carrying out multi-task target learning optimization.

The MMoE module has m different gates (gates) for m optimization objectives, i.e. n different expert networks. For example, the optimization target is a playing target and a sharing target, for example, the playing target indicates whether the user plays the item, and the sharing target indicates whether the user shares the item; then m =2,m is the same as the number of optimization targets; for example, n =4,n represents a hyper-parameter, which is generally determined in combination with off-line effects and on-line time-consuming constraints. For example, the optimization target is a playing target, a sharing target and a retention amount target, for example, the playing target indicates whether the user plays the item, the sharing target indicates whether the user shares the item, and the sharing target indicates whether the user shares the item and then brings the item into the user amount; then m =3,m is the same as the number of optimization objectives.

Wherein the fraction of the kth Gate is calculated as in the following formula (1):

g ^k(x) ＝softmax(W _gk(x) ) (1)；

wherein, g ^k(x) Is the output of the kth gate, is a softmax probability distribution; w _gk(x) Intermediate parameters representing the model structure; k (x) represents the kth gate, k corresponding to the number of optimization objectives; the MMoE module is a layer of the overall object model interface, and x refers to the input provided by the lower layer in the object model to the MMoE, for example, x is the user vector.

Wherein softmax is a normalized exponential function, or softmax function, which is a generalization of the logic function. The softmax function can compress one K-dimensional vector z containing arbitrary real numbers into another K-dimensional real vector σ (z) such that each element ranges between (0,1) and the sum of all elements is 1. The softmax function is commonly used in multi-classification problems.

Wherein, the input of tower (tower) corresponding to the kth optimization objective can be expressed as the following formula (2):

wherein k represents the kth task; n represents n expert networks; g ^k (x) _i Representing voting weights of different expert networks for an optimization goal; f. of _i (x) Representing recommendations of different expert networks for an optimization objective; the weighted sum of the two is the final proposal of a plurality of expert networks to an optimization target.

Wherein, the weight selection in the MMoE module is different, each task is provided with a Gate model, for different tasks, the output of a specific kth Gate represents the selection probability of different Expert networks Expert, a plurality of Expert networks are weighted and summed to obtain f ^k (x) And output to a specific Tower model for final output.

Wherein, the output of the k-th optimization objective is as the following formula (3):

y _k ＝ ^k (f ^k (x)) (3)；

wherein h is ^k Is the missing structure of the kth optimization target tower. Under the control of a plurality of expert networks and a plurality of gates, the optimization can be simultaneously carried outDifferent objectives, while sharing through the bottom structure, the embodiment of the present application has much smaller parameter amount compared to the separately trained network. Wherein, the bottom structure is the next layer of MMoE module layer and is used as the input of MMoE.

The embodiments of the present application provide an information recommendation method, which may be executed by a terminal or a server, or may be executed by both the terminal and the server; the embodiment of the present application is described as an example in which the information recommendation method is executed by a server.

An information recommendation method, comprising: extracting long-term interest features of the user according to the user portrait information; extracting short-term interest characteristics of the user according to click sequence information in the historical behavior data of the user; calculating an item attribute preference vector of the user according to item attribute information in the historical behavior data of the user; fusing the long-term interest characteristics, the short-term interest characteristics and the item attribute preference information of the user to obtain a user vector; optimizing a plurality of targets in a target model according to the user vector to obtain an optimized target model; and processing the item information in the item information base through the optimized target model to generate recall information corresponding to the target user, and recommending based on the recall information.

Referring to fig. 1, fig. 1 is a schematic view of an application scenario of a recommendation system provided in an embodiment of the present application, where the recommendation system includes a terminal 110 and a server 120. The terminal 110 and the server 120 are connected through a network. The terminal 110 may specifically be a desktop terminal or a mobile terminal, and the mobile terminal may specifically be at least one of a mobile phone, a tablet computer, a notebook computer, and the like. The server 120 may be implemented as a stand-alone server or a server cluster composed of a plurality of servers.

The server 120 can obtain user portrait information, click sequence information and item attribute information, extract long-term interest features of the user according to the user portrait information, extract short-term interest features of the user according to the click sequence information in the user historical behavior data, calculate item attribute preference vectors of the user according to the item attribute information in the user historical behavior data, perform fusion processing on the long-term interest features, the short-term interest features and the item attribute preference information of the user to obtain user vectors, optimize a plurality of targets in the target model according to the user vectors to obtain an optimized target model, process item information in an item information base through the optimized target model to generate recall information corresponding to the target user, and recommend the recall information based on the recall information. The server 120 sends the recall information to the terminal 110 corresponding to the target user.

It should be noted that the above application scenario is only an example, and in some embodiments, the steps of the information recommendation method may also be executed by the terminal 110. For example, the terminal 110 may directly use the configured information recommendation device to run a pre-trained target model, obtain user portrait information, click sequence information, and item attribute information through the target model, extract long-term interest features of the user according to the user portrait information, extract short-term interest features of the user according to the click sequence information in the user historical behavior data, and calculate item attribute preference vectors of the user according to the item attribute information in the user historical behavior data, then perform fusion processing on the long-term interest features, the short-term interest features, and the item attribute preference information of the user to obtain user vectors, then optimize multiple targets in the target model according to the user vectors to obtain an optimized target model, process the item information in the item information base through the optimized target model to generate recall information corresponding to the target user, and perform recommendation based on the recall information.

In some embodiments, an application (client) supporting a content push function is installed and run on the terminal 110, when the terminal 110 runs the application, a user interface for displaying push information is displayed on a screen of the terminal 110, the terminal may select recall information corresponding to a target user from item information in an item information base by performing the above-mentioned information recommendation method, push the recall information corresponding to the target user to an application program that the target user logs in, and display the recall information to the currently logged-in target user through the user interface, and the target user may browse and view the pushed information through the user interface.

For example, the recommendation system can be applied to application scenes such as text information recommendation applications, video recommendation applications, picture recommendation applications and the like.

For example, the recommendation system may be applied to a text recommendation application, that is, a recommendation text is determined from the recall text information, for example, after a user opens a recommendation APP, the terminal 110 automatically generates a request for text information recommendation for a target user, and performs processing according to user portrait information, click sequence information, and item attribute information corresponding to the target user to obtain a user vector of the target user, and determines recall text information similar to the user vector of the target user from an item information base, and then determines recommendation text information meeting the user interest preference from the recall text information, and returns the recommendation APP, so that the user obtains accurate recommendation text information.

For example, the recommendation system may be applied to a video recommendation application, and determines recommended video information from recalled video information, for example, after a user opens a recommended APP, the terminal 200 automatically generates a request for video recommendation for a target user, and performs processing according to user portrait information, click sequence information, and item attribute information corresponding to the target user to obtain a user vector of the target user, and determines recalled video information similar to the user vector of the target user from an item information base, and then determines recommended video information meeting the user interest preference from the recalled video information, and returns the recommended APP, so that the user obtains accurate recommended video information.

The following are detailed below. It should be noted that the description sequence of the following embodiments is not intended to limit the priority sequence of the embodiments.

Referring to fig. 2 to 4, fig. 2 is a schematic flow chart of an information recommendation method according to an embodiment of the present application, and fig. 3 to 4 are schematic frame diagrams of a target model according to the embodiment of the present application. The specific flow of the information recommendation method can be as follows:

and step 101, extracting long-term interest features of the user according to the user portrait information.

As shown in FIG. 3, the constructed target model is a MMoE-based multi-objective optimization recall model. The MMoE module is introduced on the basis of user multi-interest vector aggregation, and simultaneous training is carried out by optimizing a playing target (whether to play a project) and a sharing target (whether to share the project) of a user. And optimizing the multi-objective task according to the user vector.

For example, in the recommendation system of the application a, the play target prediction and the share target prediction of the user have been very important optimization targets, and therefore, the play target and the share target are selected for optimization.

First, user historical behavior data input and shared data embedding (SharedEmbeds) are performed. The User historical behavior data input includes User portrait input (User Prof Inputs), user behavior history input (Item Seq Inputs), and User behavior history Item attribute input (Item Attr Inputs). The shared data embedding comprises Item attribute embedding (Attr emails), item embedding (Item emails), user behavior embedding (Action emails) and Context embedding (Context emails). The input characteristic of the user historical behavior data input and the shared data embedding is ID information.

Then, the ID information corresponding to the embedding of the user historical behavior data input and the shared data is converted into a dense vector through an embedding Lookup Layer (variance Embedded Lookup Layer). For example, vectors are modeled using data structures of arrays, which typically store common vectors, also referred to as dense vectors.

In some embodiments, the user representation information is converted into corresponding dense vectors, and long-term interest features of the user are extracted according to the dense vectors corresponding to the user representation information.

For the long-term interest characteristics of the user, a sub-module of a user vector is modeled through a user sketch (UserProf) module and is used for representing the long-term interest characteristics of the user. The data processed by the embedding search Layer is input into a user portrait module to perform user portrait embedding (Prof fingerprints) in the user portrait module, and then is processed by a translation Layer (Transformer Layer) to obtain portrait units (Prof units) for representing long-term interest features of users.

In some embodiments, the extracting long-term interest features of the user according to the portrait information of the user includes: acquiring user portrait information, wherein the user portrait information comprises user demographic attribute information and user preference information obtained based on the user historical behavior data statistics; and extracting the long-term interest characteristics of the user according to the user population attribute information and the user preference information.

The user portrait information includes user demographic attribute information and user preference information obtained based on user historical behavior data statistics, for example, the user demographic attribute information includes gender, age, crowd and the like, and the user preference information includes processing characteristics obtained according to processing and statistics of the user historical behavior data, for example, the processing characteristics include user main interest points, tags in which the user is interested and the like.

And 102, extracting short-term interest characteristics of the user according to click sequence information in the historical behavior data of the user.

For example, taking a video item as an example, the click sequence information may include the id of the video clicked by the user, the attribute characteristics of the corresponding video, and the like.

For the short-term interest characteristics of the user, the click sequence of the user is processed through a Weighted-translation (Weighted-transformer) module in a user behavior sequence (ItemSeq) module, and potential information of the click sequence can be calculated in a parallel optimization mode. The improved of the Weighted-transform module for the transform lies in a Multi-Head Attention mechanism (Multi Head Attention) part, the model is divided into a plurality of heads (heads) to form a plurality of subspaces, and the model can pay Attention to information in different aspects by giving different weights to different heads for improvement. The Transformer is a translation module based on a self-attention mechanism, and the attention mechanism is used as a core construction of a codec to execute translation operation. For example, the Weighted-transform module may use web-based unified translation.

Specifically, data processed by an embedding search layer is input into a user behavior sequence module, and various embedded dense vectors are comprehensively processed through a sequence embedding pooling layer (Seq embedded pool) in the user behavior sequence module, wherein the sequence embedding pooling layer is a pooling layer in a transform structure; and then sequentially processing the multiple embedded dense vectors after the comprehensive processing through a Weighted-translation self-attention Layer (Weighted-transformer Layer # 1) and a Weighted-translation self-attention Layer (Weighted-transformer Layer # 2) to obtain sequence Units (Seq Units), wherein the sequence Units are used for representing short-term interest characteristics of the user.

In some embodiments, the click sequence information is converted into corresponding dense vectors, and short-term interest features of the user are extracted according to the dense vectors corresponding to the click sequence information.

Step 103, calculating an item attribute preference vector of the user according to the item attribute information in the user historical behavior data.

The item attribute information comprises item labels, item categories, item title word segments, item themes and the like. Taking a video project as an example, the project attribute information includes a video tag, a video category, a video title word, a video theme, and the like.

In some embodiments, the calculating an item attribute preference vector of a user according to item attribute information in the user historical behavior data includes: fusing the long-term interest features and the short-term interest features of the user to obtain a first fusion vector; and processing the first fusion vector and the item attribute information to calculate an item attribute preference vector of the user.

For the Item attribute preference vector of the user, an Item attribute preference module is designed, for example, named as Attr2Item module, a first fusion vector (long-short term interest vector of the user) is obtained by fusing long term interest features and short term interest features of the user, the first fusion vector is input into the Attr2Item module, and the first fusion vector and the Item attribute information are processed by using a multi-head Attention mechanism (Attention) in the Attr2Item module, so as to calculate the Item attribute preference vector of the user for the Item. The Attr2Item module is used for modeling the relationship between the properties of the user history interactive Item and the recommendation structure. Such as which tags the user is primarily interested in, which tags are used for recommendations. And the Attr2Item structure integrates how to obtain which tags the user is interested in and how to establish an association with the recommendation structure. Specifically, a long-term interest feature of a user output by a user portrait module and a short-term interest feature of the user output by a user click sequence module are fused (Aggregate) to obtain a first fusion vector (a long-term and short-term interest vector of the user), the first fusion vector is input into a topN self-attention layer (topN attention layer) in an Attr2Item module, item attribute information (Item attribute embedding) in shared data embedding is input into the topN self-attention layer, and the first fusion vector and the Item attribute information are processed through the topN self-attention layer to calculate an Item attribute preference vector of the user for an Item. Wherein topN self-attention layer is used to focus attention on the top N data.

In some embodiments, the item attribute information is converted into corresponding dense vectors, and the item attribute preference vector of the user is calculated according to the dense vectors corresponding to the item attribute information.

And 104, fusing the long-term interest characteristics, the short-term interest characteristics and the item attribute preference information of the user to obtain a user vector.

For example, the vectors of the long-term interest features, the short-term interest features and the item attribute preference information of the user are added and aggregated into a final expression vector of the user, that is, the user vector is obtained.

And 105, optimizing a plurality of targets in the target model according to the user vector to obtain an optimized target model.

In some embodiments, the optimizing the multiple targets in the target model according to the user vector to obtain an optimized target model includes:

and optimizing the playing target and the sharing target in a target model according to the user vector to obtain an optimized target model.

In some embodiments, the optimizing the playing objective and the sharing objective in the objective model according to the user vector to obtain an optimized objective model includes:

processing the user vector through the multitask learning module to split the user vector into common information and difference information, wherein the common information is the common information corresponding to the playing target and the sharing target, and the difference information is the difference information corresponding to the playing target and the sharing target;

and optimizing the playing target and the sharing target in the target model according to the common information and the difference information to obtain an optimized target model.

For example, a shared multi-task learning (SharedMMoE) module performs optimization of multi-objective tasks based on user vectors. Specifically, the user vector is processed through a shared MMoE Layer (shared MMoE Layer) in the SharedMMoE module to achieve optimization of a Share target (Share Unit) and a Play target (Play Unit). Different vectors are then generated for each target and the generated vectors are combined (Concat). In fig. 3, the structures of extracting and fusing the long-term interest features, the short-term interest features and the attribute preference vectors to the user vector are unified, then the obtained user vector is split into data required by two optimization targets through a SharedMMoE module, a plurality of expert networks work in cooperation during data splitting, and the SharedMMoE module uses the user vector which is uniformly input by a lower layer.

And calculating the Cosine similarity between the user vector of the sample user and the item vector related in the user behavior history of the sample user based on an Online service (Online Serving) to obtain a candidate recall of the sample user U, namely calculating the top K items (topK) most interested in by the sample user. In addition, based on Offline service (Offline Serving), the Play Target embedding (Play Target embedding), the Share Target embedding (Share Target embedding) and the result of the shared mmoe module after processing based on the Share Target and the Play Target are input into the overall Loss function (Sampled Loss) of the Target model for training, so that when the first K items in which the sample user is most interested are trained to be most matched with the historical recommendation items, synchronous optimization of multiple targets in the Target model is completed, and the optimized Target model is obtained.

As shown in fig. 3, after the user vectors are aggregated, the user's playing goal and sharing goal for the project are optimized respectively through the MMoE module. The embodiment of the application further improves the MMoE module, different information can be extracted from different expert networks in the MMoE module, and the input of each target Tower is controlled through different gates. In a target recommendation scene, since the click rate and the share rate of the user are highly correlated, the embodiment of the application considers that different expert networks may have common interests, so a gateway shared structure is designed in the MMoE module to extract the common information, and the gateway shared structure is used for extracting only supplementary information of different expert network information for different gates. Wherein the basic MMoE module has the same input for multiple expert networks. The embodiment of the application takes the difference of different targets into consideration, introduces a common structure (such as a Gateshared structure) for modeling common information (information of the same or similar parts) in a plurality of targets, and performs explicit distinction treatment by splitting original input information into common input information and different input information which is different along with the targets.

In the embodiment of the application, the sequence in each user play history (session) in the historical behavior data of each user is selected as the input data of the ItemSeq structure, and because the time difference of the information in the user play histories of the users is not large, the recommendation result obtained by using the training data can obtain the feedback of the users in a shorter time interval. On the contrary, if the time span of the used training data is too large, the real-time performance of the recommended result is not high, and the short-term interest of the user is not easy to grasp. For example, if the historical length of the training data can be expanded, the user's long-term interest can be compromised with the short-term data, such as to compromise the effectiveness and on-line pattern constraints, the historical length can be set to 64. The ItemSeq structure is used for acquiring a user behavior sequence corresponding to the Item in the user play history, and a storage module in the recommendation system stores the history behavior (such as the user play history) of each user.

The training data includes user portrait information, user historical behavior attribute information, target item information, behavior information (such as sharing and playing) between the user and the target item, and the like.

After the training data is encoded by using the Transformer-based multi-objective optimization network, the user vector E is required to be used _u To predict the outcome that needs to be recommended. Suppose the project candidate pool is [ X1, X2, X3, … …, X _N ]Calculating the Embedding (Embedding) vector and the user vector E of each candidate item _u Given two vectors, the Cosine similarity of the two vectors is calculated as the following formula (4), and then the vectors with the highest Cosine similarity are recommended:

wherein u represents a user vector; v denotes the item vector, i.e. the embedding vector for each candidate.

For a user U, fusing user portrait information, user click sequence information and project attribute information to obtain a user vector E _u Then calculates the user vector E _u And (3) calculating two values of the Cosine similarity score of each candidate item, wherein one value is a value corresponding to a sharing target (whether to share an item), and the other value is a value corresponding to a playing target (whether to play the item), and under a target recommendation scene, a weight is given to the sharing target and the estimated value corresponding to the click target, and then the final score is obtained through weighting. For example, different targets correspondingly pull items and scores of different topns, and a final score of the items is obtained by integrating multiple targets, so as to give a recommendation candidate set. Assuming that only 10 candidate pools exist, the sum-user vector E is obtained after calculation _u Most similar candidate itemX3, X5 and X7, the most similar candidates (X3, X5 and X7) are taken as recall candidates, so that the recommendation system completes the candidate recall for the user U.

In some embodiments, the plurality of targets further includes a retention amount target, the method further comprising:

acquiring project feedback information containing the remaining amount;

and optimizing the playing target, the sharing target and the retention amount target in a target model according to the user vector and the project feedback information to obtain an optimized target model.

In some embodiments, optimizing the playing objective, the sharing objective, and the retention amount objective in an objective model according to the user vector and the project feedback information to obtain an optimized objective model includes: optimizing the playing target in a target model according to the user vector; optimizing the sharing target and the retention amount target in a target model according to the user vector and the project feedback information; and obtaining an optimized target model according to the optimized playing target, the sharing target and the retention amount target.

The remaining amount in the item feedback information may affect the sharing target. But does not affect the target of the playback. For example, when the index of the user sharing the target is modeled, considering that most of videos shared by the user in a target recommendation scene are new users, the amount of the retained videos received by the new users needs to be modeled. Therefore, the reserve amount of the video shared by the other people received by the user is used as an additional modeling target and can be called a reserve amount target, and the reserve amount target is used as a weight in multiple targets. Related to the target recommendation scene, the target recommendation scene attaches importance to the number of Active users (DAU) on a day, and after the project is shared, the User retains the project and brings access to the User, so that the DAU is increased. An additional objective is the amount of users to be brought in after each sharing, and the whole recommendation system tends to preferentially select candidates with large amounts of users to be recommended under the same condition. The amount of remaining may represent the amount brought to the user after sharing.

Therefore, the reserve amount is used as a new target, and the previous sharing target and the playing target are optimized together: on one hand: all targets share bottom-level input; on the other hand: the remaining amount (the amount of users brought in after sharing) helps the recommendation system make a preferential choice to select a candidate item that can promote the DAU.

The multi-objective optimization model is observed from the perspective of a real scene, if the retention amount of a new user is not modeled, and only the sharing target and the playing target of the user are optimized, the new user does not return visit within a few days after the user shares videos with other users, which is also equivalent to that the optimization target of the sharing target does not bring good income to the whole recommendation system, but rather loses a part of potential long-term users. Therefore, while optimizing the playing goal and the sharing goal, the retention amount goal needs to be introduced to be optimized together.

As shown in fig. 4, based on fig. 3, item feedback information (Item feeds) including the remaining amount (amount of Shared user) is introduced, and then the user vector is processed by the MMoE Layer (wmted Shared MMoE Layer) in the wmt Shared MMoE module to optimize the sharing target (Share Unit), the playing target (Play Unit), and the remaining amount target (Wgt Unit). And calculating the Cosine similarity between the user vector of the sample user and the item vector related in the user behavior history of the sample user based on an Online service (Online Serving) to obtain a candidate recall of the sample user U, namely calculating the top K items (topK) most interested in by the sample user. In addition, based on Offline service (Offline Serving), the playing Target embedding (Play Target embedding), the sharing Target embedding (Share Target embedding) and the result of the WgtSharedMMoE module after being processed based on the sharing Target and the playing Target are input into the overall Loss function (Sampled Loss) of the Target model for training, and the result of the WgtSharedMMoE module after being processed based on the sharing Target and the playing Target is input into the reserve amount Target Loss function (Wgt Loss) for training, so that when the first K items which are most interesting to the sample user are trained to be most matched with the historical recommended items, synchronous optimization of a plurality of targets (the playing Target, the sharing Target and the reserve amount Target) in the Target model is completed, and the optimized Target model is obtained.

And 106, processing the item information in the item information base through the optimized target model to generate recall information corresponding to the target user, and recommending based on the recall information.

In some embodiments, the processing, by the optimized target model, the item information in the item information base to generate recall information corresponding to a target user, and recommending based on the recall information includes: obtaining a user vector of a target user through the optimized target model; calculating cosine similarity between the user vector of the target user and the project vector in the project information of the project information base; generating recall information corresponding to the target user according to the cosine similarity; and determining recommendation information from the recall information, and recommending the recommendation information to the target user.

For example, user portrait information, click sequence information and item attribute information corresponding to a target user are processed through an optimized target model to obtain a user vector of the target user, cosine similarity between the user vector of the target user and item vectors in item information of an item information base is calculated, recall information corresponding to the target user is generated according to the cosine similarity between the user vector of the target user and the item vectors in the item information base, recommendation information is determined from the recall information, and the recommendation information is recommended to the target user. The target user is a user who uses the recommendation system currently or a currently logged-in user who logs in an application APP with the recommendation system currently.

For example, when the user vector of the target user needs to be used to predict the result that needs to be recommended, for example, the candidate pool in the project information base is [ Y1, Y2, Y3, … …, Y _N ]Calculating the Cosine similarity between the item vector of each candidate item in the item information base and the user vector, and given two vectors, the Cosine similarity calculation of the two vectors can refer to formula (4),are not described in detail herein; then, recommending the vectors with the highest Cosine similarity. For the target user, user portrait information, click sequence information and item attribute information of the target user are fused to obtain a user vector of the target user, then a Cosine similarity score of the user vector and the item vector of each candidate item is calculated, the candidate items most similar to the target user vector are obtained after calculation and are Y1, Y2 and Y5, and the most similar candidate items (Y1, Y2 and Y5) are used as recall candidate items (recall information of the target user) so that the recommendation system can complete candidate item recall for the target user U.

In some embodiments, the generating recall information corresponding to a target user according to the cosine similarity includes: determining a first candidate item set from item information in the item information base according to the cosine similarity; performing initial selection on the first candidate item set to obtain a second candidate item set, wherein the number of items of the second candidate item set is smaller than that of the first candidate item set; and sorting the candidate items in the second candidate item set, and selecting K candidate items before ranking from the sorted second candidate item set as the recall information corresponding to the target user.

For example, a recommendation system corresponding to a target application scenario may include a recall logic unit, a primary selection logic unit, and a Rank logic unit. The recall logic unit is used for pulling (recalling) data (Document) according to dimensions such as various precise individualization, general individualization and heat according to the image information of a specific user. The primary selection logic unit is used for primarily screening a large number of recall results according to a specific primary selection rule (for example, the specific primary selection rule is set according to parameters such as user document correlation, timeliness, regionality and diversity) so as to reduce the calculation scale of the ranking logic unit, for example, tens of thousands of items are pulled when the recall logic unit recalls, and then data screening is carried out through the primary selection logic unit so that tens of thousands of items are changed into thousands of items. And the sequencing logic unit is used for sequencing the final result output by the target model and recommending the final result to the user.

For example, a first candidate item set is determined from item information in the item information base according to cosine similarity between a user vector of a target user and item vectors in the item information base. For example, the recall logic unit may calculate a plurality of pre-estimated values based on the MMoE multitask learning model, and then use the plurality of pre-estimated values to find the TopK candidate items of most interest to the user, and use the TopK candidate items as the first candidate item set. The candidate logic in the recall is to store each item vector offline and generate the user vector online, since the context of the user online is dynamic, so the recommendation system must keep the user vector generated online in order to capture the user's interest preferences at the current time.

For example, the first candidate item set is initially selected based on a specific initial selection rule to obtain a second candidate item set, where the number of items of the second candidate item set is smaller than the number of items of the first candidate item set, for example, the specific initial selection rule is set according to parameters such as user document relevance, timeliness, locality, and diversity. For example, in the primary selection logic unit, the MMoE multitask learning model may be used not only to help the candidate screening, but also in the ranking stage, the user vector of each user is already calculated in the recall stage, and in order to calculate the Cosine similarity between the embedded (Embedding) vector of the candidate and the user vector of the target user in the primary selection stage, the item with higher score is screened out and input to the ranking logic unit for processing.

For example, the candidates in the second candidate item set are ranked, and K top-ranked candidate items are selected from the ranked second candidate item set as the recall information corresponding to the target user. For example, the final result output by the target model is sorted by the sorting logic unit, and the items with K top ranks are recommended to the target user.

The items in the recall information can be all recommended to the target user, and the items with the top rank can also be recommended to the target user. That is, all items in the recall information may be determined as recommendation information, or an item ranked before a preset ranking in the recall information may be determined as recommendation information, and the recommendation information is recommended to the target user.

All the above technical solutions can be combined arbitrarily to form the optional embodiments of the present application, and are not described herein again.

According to the method and the device, long-term interest characteristics of the user are extracted according to the portrait information of the user; extracting short-term interest characteristics of the user according to click sequence information in the historical behavior data of the user; calculating an item attribute preference vector of the user according to item attribute information in the historical behavior data of the user; fusing the long-term interest characteristics, the short-term interest characteristics and the item attribute preference information of the user to obtain a user vector; optimizing a plurality of targets in the target model according to the user vector to obtain an optimized target model; and processing the item information in the item information base through the optimized target model to generate recall information corresponding to the target user, and recommending based on the recall information. According to the method and the device, the user portrait information, the click sequence information and the project attribute information are aggregated into the user vector representing the interest preference of the user, and the multiple targets in the target model are optimized simultaneously based on the user vector, so that the interest preference of the user can be reflected by the model, and the recommendation accuracy of the model is improved.

In order to better implement the information recommendation method in the embodiment of the present application, an embodiment of the present application further provides an information recommendation device. Referring to fig. 5, fig. 5 is a schematic structural diagram of an information recommendation device according to an embodiment of the present application. The information recommendation apparatus 500 may include:

a first extraction unit 501, configured to extract a long-term interest feature of a user according to user portrait information;

the second extraction unit 502 is configured to extract short-term interest features of the user according to click sequence information in the user historical behavior data;

a calculating unit 503, configured to calculate an item attribute preference vector of the user according to item attribute information in the user historical behavior data;

a fusion unit 504, configured to perform fusion processing on the long-term interest feature, the short-term interest feature, and the item attribute preference information of the user to obtain a user vector;

an optimizing unit 505, configured to optimize multiple targets in a target model according to the user vector to obtain an optimized target model;

a recommending unit 506, configured to process the item information in the item information base through the optimized target model to generate recall information corresponding to the target user, and recommend based on the recall information.

In some embodiments, the first extraction unit 501 is configured to: acquiring user portrait information, wherein the user portrait information comprises user population attribute information and user preference information obtained based on the user historical behavior data statistics; and extracting the long-term interest characteristics of the user according to the user population attribute information and the user preference information.

In some embodiments, the calculating unit 503 is configured to: fusing the long-term interest features and the short-term interest features of the user to obtain a first fusion vector; and processing the first fusion vector and the item attribute information to calculate an item attribute preference vector of the user.

In some embodiments, the multiple targets include a playing target and a sharing target, and the optimizing unit 505 is configured to optimize the playing target and the sharing target in a target model according to the user vector to obtain an optimized target model.

In some embodiments, the objective model comprises a multitask learning module, and the optimization unit 505 is configured to: processing the user vector through the multitask learning module to split the user vector into common information and difference information, wherein the common information is the common information corresponding to the playing target and the sharing target, and the difference information is the difference information corresponding to the playing target and the sharing target; and optimizing the playing target and the sharing target in the target model according to the common information and the difference information to obtain an optimized target model.

In some embodiments, the plurality of objectives further comprises a retention amount objective, the optimization unit 505 is configured to: acquiring project feedback information containing the reserved quantity; and optimizing the playing target, the sharing target and the retention amount target in a target model according to the user vector and the project feedback information to obtain an optimized target model.

In some embodiments, the optimizing unit 505 is configured to: optimizing the playing target in a target model according to the user vector; optimizing the sharing target and the retention amount target in a target model according to the user vector and the project feedback information; and obtaining an optimized target model according to the optimized playing target, the sharing target and the retention amount target.

In some embodiments, the recommending unit 506 is configured to: obtaining a user vector of a target user through the optimized target model; calculating cosine similarity between the user vector of the target user and the project vector in the project information of the project information base; generating recall information corresponding to the target user according to the cosine similarity; and determining recommendation information from the recall information, and recommending the recommendation information to the target user.

In some embodiments, the recommending unit 506 is configured to generate recall information corresponding to the target user according to the cosine similarity, and specifically includes: determining a first candidate item set from item information in the item information base according to the cosine similarity; performing initial selection on the first candidate item set to obtain a second candidate item set, wherein the number of items of the second candidate item set is smaller than that of the first candidate item set; and sorting the candidate items in the second candidate item set, and selecting K candidate items before ranking from the sorted second candidate item set as the recall information corresponding to the target user.

In some embodiments, the first extracting unit 501 is configured to convert the user portrait information into corresponding dense vectors, and extract long-term interest features of the user according to the corresponding dense vectors of the user portrait information;

the second extracting unit 502 is configured to convert the click sequence information into corresponding dense vectors, and extract short-term interest features of the user according to the dense vectors corresponding to the click sequence information;

the calculating unit 503 is configured to convert the item attribute information into a corresponding dense vector, and calculate an item attribute preference vector of the user according to the dense vector corresponding to the item attribute information.

In the information recommendation apparatus 500 provided in the embodiment of the application, the first extraction unit 501 extracts long-term interest features of a user according to user image information; the second extraction unit 502 extracts the short-term interest features of the user according to the click sequence information in the user historical behavior data; the calculating unit 503 calculates an item attribute preference vector of the user according to item attribute information in the user historical behavior data; the fusion unit 504 performs fusion processing on the long-term interest features, the short-term interest features and the item attribute preference information of the user to obtain a user vector; the optimization unit 505 optimizes a plurality of targets in the target model according to the user vector to obtain an optimized target model; the recommending unit 506 processes the item information in the item information base through the optimized target model to generate recall information corresponding to the target user, and recommends based on the recall information. According to the method and the device, the user portrait information, the click sequence information and the project attribute information are aggregated into the user vector representing the interest preference of the user, and the multiple targets in the target model are optimized simultaneously based on the user vector, so that the interest preference of the user can be reflected by the model, and the recommendation accuracy of the model is improved.

Correspondingly, the embodiment of the application also provides a computer device, and the computer device can be a terminal or a server. As shown in fig. 6, the computer device may include Radio Frequency (RF) circuitry 601, memory 602 including one or more computer-readable storage media, input unit 603, display unit 604, sensor 605, audio circuitry 606, wireless Fidelity (WiFi) module 607, processor 608 including one or more processing cores, and power supply 609. Those skilled in the art will appreciate that the computer device configuration illustrated in FIG. 6 does not constitute a limitation of computer devices, and may include more or fewer components than those illustrated, or some components may be combined, or a different arrangement of components. Wherein:

the RF circuit 601 may be used for receiving and transmitting signals during a message transmission or communication process, and in particular, for receiving downlink messages from a base station and then processing the received downlink messages by one or more processors 608; in addition, data relating to uplink is transmitted to the base station. In addition, the RF circuit 601 may also communicate with networks and other devices via wireless communications.

The memory 602 may be used to store software programs and modules, and the processor 608 executes various functional applications and data processing by operating the software programs and modules stored in the memory 602. The memory 602 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application program required for at least one function, and the like; the storage data area may store data created according to use of the computer device, and the like.

The input unit 603 may be used to receive input numeric or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

The display unit 604 may be used to display information input by or provided to a user as well as various graphical user interfaces of the computer device, which may be made up of graphics, text, icons, video, and any combination thereof. The display unit 604 may include a display panel.

The computer device may also include at least one sensor 605, such as a light sensor, motion sensor, and other sensors.

Audio circuitry 606, speakers, and microphones may provide an audio interface between a user and a computer device. The audio circuit 606 may transmit the electrical signal converted from the received audio data to a speaker, and convert the electrical signal into a sound signal for output; on the other hand, the microphone converts the collected sound signals into electrical signals, which are received by the audio circuit 606 and converted into audio data, which are then processed by the audio data output processor 608, either through the RF circuit 601 for transmission to, for example, another computer device, or output to the memory 602 for further processing. Audio circuitry 606 may also include an earbud jack to provide communication of peripheral headphones with the computer device.

WiFi belongs to short-range wireless transmission technology, and the computer device can help the user send and receive e-mail, browse web pages, access streaming media, etc. through the WiFi module 607, which provides wireless broadband internet access for the user. Although fig. 6 shows the WiFi module 607, it is understood that it does not necessarily constitute a computer device, and may be omitted as needed within a scope not changing the essence of the invention.

The processor 608 is a control center of the computer device, connects various parts of the entire cellular phone using various interfaces and lines, performs various functions of the computer device and processes data by operating or executing software programs and/or modules stored in the memory 602 and calling data stored in the memory 602, thereby monitoring the computer device as a whole.

The computer device also includes a power supply 609 (e.g., a battery) for powering the various components, which may be logically coupled to the processor 608 via a power management system that provides management of charging, discharging, and power consumption.

Although not shown, the computer device may further include a camera, a bluetooth module, etc., which will not be described herein. Specifically, in this embodiment, the processor 608 in the computer device loads the executable file corresponding to one or more processes of the computer program into the memory 602 according to the following instructions, and the processor 608 runs the computer program stored in the memory 602, so as to implement various functions:

extracting long-term interest characteristics of the user according to the user portrait information; extracting short-term interest characteristics of the user according to click sequence information in the historical behavior data of the user; calculating an item attribute preference vector of the user according to item attribute information in the historical behavior data of the user; fusing the long-term interest characteristics, the short-term interest characteristics and the item attribute preference information of the user to obtain a user vector; optimizing a plurality of targets in a target model according to the user vector to obtain an optimized target model; and processing the item information in the item information base through the optimized target model to generate recall information corresponding to the target user, and recommending based on the recall information.

The above operations can be implemented in the foregoing embodiments, and are not described in detail herein.

It will be understood by those skilled in the art that all or part of the steps of the methods of the above embodiments may be performed by instructions or by associated hardware controlled by the instructions, which may be stored in a computer readable storage medium and loaded and executed by a processor.

The embodiment of the application also provides a computer readable storage medium for storing the computer program. The computer-readable storage medium can be applied to a computer device, and the computer program enables the computer device to execute the corresponding process in the information recommendation method in the embodiment of the present application, which is not described herein again for brevity.

Embodiments of the present application also provide a computer program product including computer instructions stored in a computer-readable storage medium. The processor of the computer device reads the computer instruction from the computer-readable storage medium, and executes the computer instruction, so that the computer device executes a corresponding process in the information recommendation method in the embodiment of the present application, which is not described herein again for brevity.

Embodiments of the present application also provide a computer program, which includes computer instructions stored in a computer-readable storage medium. The processor of the computer device reads the computer instruction from the computer-readable storage medium, and executes the computer instruction, so that the computer device executes the corresponding process in the information recommendation method in the embodiment of the present application, which is not described herein again for brevity.

It should be understood that the processor of the embodiments of the present application may be an integrated circuit chip having signal processing capabilities. In implementation, the steps of the above method embodiments may be performed by integrated logic circuits of hardware in a processor or by instructions in the form of software. The Processor may be a general purpose Processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), an off-the-shelf Programmable Gate Array (FPGA) or other Programmable logic device, discrete Gate or transistor logic device, or discrete hardware components. The various methods, steps, and logic blocks disclosed in the embodiments of the present application may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of the method disclosed in connection with the embodiments of the present application may be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software module may be located in ram, flash memory, rom, prom, or eprom, registers, etc. storage media as is well known in the art. The storage medium is located in a memory, and a processor reads information in the memory and completes the steps of the method in combination with hardware of the processor.

It will be appreciated that the memory in the embodiments of the subject application can be either volatile memory or nonvolatile memory, or can include both volatile and nonvolatile memory. The non-volatile Memory may be a Read-Only Memory (ROM), a Programmable ROM (PROM), an Erasable PROM (EPROM), an Electrically Erasable PROM (EEPROM), or a flash Memory. The volatile Memory may be a Random Access Memory (RAM) which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static random access memory (Static RAM, SRAM), dynamic Random Access Memory (DRAM), synchronous Dynamic random access memory (Synchronous DRAM, SDRAM), double Data Rate Synchronous Dynamic random access memory (DDR SDRAM), enhanced Synchronous SDRAM (ESDRAM), synchronous link SDRAM (SLDRAM), and Direct Rambus RAM (DR RAM). It should be noted that the memory of the systems and methods described herein is intended to comprise, without being limited to, these and any other suitable types of memory.

It should be understood that the above memories are exemplary but not limiting illustrations, for example, the memories in the embodiments of the present application may also be Static Random Access Memory (SRAM), dynamic random access memory (dynamic RAM, DRAM), synchronous Dynamic Random Access Memory (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (enhanced SDRAM, ESDRAM), synchronous Link DRAM (SLDRAM), direct Rambus RAM (DR RAM), and the like. That is, the memory in the embodiments of the present application is intended to comprise, without being limited to, these and any other suitable types of memory.

Those of ordinary skill in the art will appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware or combinations of computer software and electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.

It is clear to those skilled in the art that, for convenience and brevity of description, the specific working processes of the above-described systems, apparatuses and units may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again.

In the several embodiments provided in the present application, it should be understood that the disclosed system, apparatus and method may be implemented in other ways. For example, the above-described apparatus embodiments are merely illustrative, and for example, the division of the units is only one logical division, and other divisions may be realized in practice, for example, a plurality of units or components may be combined or integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, devices or units, and may be in an electrical, mechanical or other form.

The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.

In addition, functional units in the embodiments of the present application may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit.

The functions, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present application or portions thereof that substantially contribute to the prior art may be embodied in the form of a software product stored in a storage medium and including instructions for causing a computer device (which may be a personal computer or a server) to execute all or part of the steps of the method according to the embodiments of the present application. And the aforementioned storage medium includes: a U disk, a removable hard disk, a ROM, a RAM, a magnetic disk or an optical disk, and various media capable of storing program codes.

The above description is only for the specific embodiments of the present application, but the scope of the present application is not limited thereto, and any person skilled in the art can easily think of the changes or substitutions within the technical scope of the present application, and shall be covered by the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. An information recommendation method, characterized in that the method comprises:

extracting long-term interest characteristics of the user according to the user portrait information;

2. The information recommendation method of claim 1, wherein the extracting long-term interest features of the user according to the portrait information of the user comprises:

acquiring user portrait information, wherein the user portrait information comprises user population attribute information and user preference information obtained based on the user historical behavior data statistics;

and extracting the long-term interest characteristics of the user according to the user population attribute information and the user preference information.

3. The information recommendation method of claim 1, wherein said calculating an item attribute preference vector for a user based on item attribute information in said user historical behavior data comprises:

fusing the long-term interest features and the short-term interest features of the user to obtain a first fusion vector;

and processing the first fusion vector and the item attribute information to calculate an item attribute preference vector of the user.

4. The information recommendation method of claim 1, wherein the plurality of objectives include a play objective and a share objective, and the optimizing the plurality of objectives in the objective model according to the user vector to obtain an optimized objective model comprises:

5. The information recommendation method according to claim 4, wherein the objective model comprises a multitask learning module, and the optimizing the playing objective and the sharing objective in the objective model according to the user vector to obtain an optimized objective model comprises:

6. The information recommendation method of claim 4, wherein the plurality of objectives further include a retention amount objective, the method further comprising:

acquiring project feedback information containing the reserved quantity;

7. The information recommendation method according to claim 6, wherein optimizing the playing goal, the sharing goal, and the retention amount goal in a goal model according to the user vector and the item feedback information to obtain an optimized goal model comprises:

optimizing the playing target in a target model according to the user vector;

optimizing the sharing target and the retention amount target in a target model according to the user vector and the project feedback information;

and obtaining an optimized target model according to the optimized playing target, the sharing target and the retention amount target.

8. The information recommendation method of claim 1, wherein the processing item information in an item information base through the optimized target model to generate recall information corresponding to a target user and making a recommendation based on the recall information comprises:

obtaining a user vector of a target user through the optimized target model;

calculating cosine similarity between the user vector of the target user and the project vector in the project information of the project information base;

generating recall information corresponding to the target user according to the cosine similarity;

and determining recommendation information from the recall information, and recommending the recommendation information to the target user.

9. The information recommendation method according to claim 8, wherein the generating recall information corresponding to the target user according to the cosine similarity comprises:

determining a first candidate item set from item information in the item information base according to the cosine similarity;

performing initial selection on the first candidate item set to obtain a second candidate item set, wherein the number of items of the second candidate item set is smaller than that of the first candidate item set;

and sorting the candidate items in the second candidate item set, and selecting K candidate items before ranking from the sorted second candidate item set as the recall information corresponding to the target user.

10. The information recommendation method of claim 1, further comprising;

converting the user portrait information into corresponding dense vectors, and extracting long-term interest features of the user according to the dense vectors corresponding to the user portrait information;

converting the click sequence information into corresponding dense vectors, and extracting short-term interest features of the user according to the dense vectors corresponding to the click sequence information;

and converting the item attribute information into corresponding dense vectors, and calculating the item attribute preference vectors of the users according to the dense vectors corresponding to the item attribute information.

11. An information recommendation apparatus, characterized in that the apparatus comprises:

the first extraction unit is used for extracting long-term interest features of the user according to the user portrait information;

12. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program adapted to be loaded by a processor for performing the steps of the information recommendation method according to any one of claims 1-10.

13. A computer device, characterized in that the computer device comprises a processor and a memory, wherein the memory stores a computer program, and the processor is used for executing the steps in the information recommendation method according to any one of claims 1-10 by calling the computer program stored in the memory.