CN114996561B

CN114996561B - Information recommendation method and device based on artificial intelligence

Info

Publication number: CN114996561B
Application number: CN202110231593.2A
Authority: CN
Inventors: 徐程程
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2021-03-02
Filing date: 2021-03-02
Publication date: 2024-03-29
Anticipated expiration: 2041-03-02
Also published as: CN114996561A

Abstract

The application provides an information recommendation method and device based on artificial intelligence; to artificial intelligence techniques, the method comprising: extracting user behavior characteristics from behavior data of a plurality of interactive media accounts corresponding to the user; coding processing is carried out based on the user behavior characteristics to obtain user characteristic vectors; determining a plurality of recall media accounts satisfying similar conditions with the user feature vector; and generating information to be recommended based on the plurality of recall media accounts, and executing recommendation operation corresponding to the user based on the information to be recommended. Through the method and the device, the user interests can be fully mined to improve the recommendation accuracy.

Description

Information recommendation method and device based on artificial intelligence

Technical Field

The application relates to an artificial intelligence technology, in particular to an information recommendation method and device based on artificial intelligence.

Background

Artificial intelligence (AI, artificial Intelligence) is the theory, method and 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.

Information recommendation is an important application of artificial intelligence, and various strategies and models are usually processed in parallel in the recall process in a recommendation system, for example, related information (also called materials, articles, videos and the like) is searched based on user portraits, similar information is searched based on the user recently clicking, popular information is searched and the like. In most information flow scenes, different types of information are displayed in a mixed mode, such as articles and videos are cross-appearing, and the recall algorithm in the related art can recall different information respectively, namely, the recall complexity is greatly increased when the information is faced with multiple types of information, interests of users in different types of information cannot be fused, so that the recommended information cannot meet the abundant interests of the users, and bad experience is caused to the users.

Therefore, there is a lack of effective schemes for fusing various types of information of interest to a user for accurate recommendation in the related art.

Disclosure of Invention

The embodiment of the application provides an information recommending method, device, electronic equipment and computer readable storage medium based on artificial intelligence, which can fully mine user interests to improve recommending accuracy.

The technical scheme of the embodiment of the application is realized as follows:

the embodiment of the application provides an information recommendation method based on artificial intelligence, which comprises the following steps:

extracting user behavior characteristics from behavior data of a plurality of interactive media accounts corresponding to the user; coding processing is carried out based on the user behavior characteristics to obtain user characteristic vectors; determining a plurality of recall media accounts satisfying similar conditions with the user feature vector; and generating information to be recommended based on the plurality of recall media accounts, and executing recommendation operation corresponding to the user based on the information to be recommended.

The embodiment of the application provides an information recommendation device based on artificial intelligence, which comprises:

the extraction module is used for extracting user behavior characteristics from behavior data of the user corresponding to the plurality of interactive media account numbers; the coding module is used for carrying out coding processing based on the user behavior characteristics to obtain user characteristic vectors; a recall module for determining a plurality of recall media accounts satisfying a similar condition to the user feature vector; and the recommending module is used for generating information to be recommended based on the plurality of recall media accounts and executing recommending operation corresponding to the user based on the information to be recommended.

In the above scheme, the extracting module is further configured to map behavior data of the user for a plurality of information into behavior data of the user for an interactive media account for publishing the information; wherein the behavior data of the user for a plurality of information characterizes at least one of the following behaviors: and the user focuses on the behavior of the information published by the interactive media account and the behavior of subscribing the information published by the interactive media account.

In the above scheme, the coding module is further configured to determine a user collaborative filtering feature and a user graph feature, sequentially connect the user collaborative filtering feature, the user graph feature, and the user behavior feature obtained by extraction, and use a new user behavior feature obtained by connection as a user behavior feature for performing the coding processing; wherein the user collaborative filtering feature is determined by decomposing a matrix comprising the behavior data, and the user graph feature is determined based on mapping a directed graph comprising the behavior data.

In the above scheme, the encoding module is further configured to extract a user attribute feature from the attribute data of the user; rectifying the user behavior characteristics and the user attribute characteristics, or rectifying the user behavior characteristics, normalizing the obtained rectifying result, and taking the obtained normalizing result as a user characteristic vector obtained by encoding; wherein the user behavior feature comprises at least one of: the method comprises the steps of enabling the user to interact with a media account number, enabling the user to interact with the media account number, and enabling the user to interact with the media account number.

In the above scheme, the recall module is further configured to obtain media account feature vectors of a plurality of candidate media accounts, determine cosine distances between the media account feature vectors of the plurality of candidate media accounts and the user feature vectors, and use the cosine distances as the similarity; and determining a plurality of candidate media accounts with similarity exceeding a similarity threshold value with the user feature vector as a plurality of recall media accounts meeting the similarity condition.

In the above solution, the recall module is further configured to perform, for each candidate media account of the plurality of candidate media accounts, the following processing: extracting media account characteristics from attribute data of the candidate media account; rectifying the media account characteristics, normalizing the obtained rectifying result, and taking the obtained normalizing result as a media account characteristic vector; wherein the media account characteristics include at least one of: the label characteristics of the media account, the channel characteristics of the media account and the vermicelli quantity characteristics of the media account.

In the above scheme, the recall module is further configured to determine a media account collaborative filtering feature and a media account map feature of the candidate media account, connect the media account collaborative filtering feature, the media account map feature and the media account feature obtained by extraction, and use a new media account feature obtained by connection as a media account feature for performing the rectification processing; wherein the media account collaborative filtering feature is determined by decomposing a matrix comprising the behavior data, and the media account graph feature is determined based on mapping a directed graph comprising the behavior data.

In the above solution, the recommendation module is further configured to generate the information to be recommended by at least one of the following means: generating information to be recommended for recommending the media account based on the plurality of recall media accounts; generating information to be recommended for recommending the information published by the media account based on the information published by the plurality of recall media accounts; predicting to obtain a score of the information to be recommended based on the information feature vector of the information to be recommended, the user feature vector of the user and the cross feature vector of the user and the information to be recommended, wherein the score characterizes the similarity of the information to be recommended and the user; sorting the candidate information to be recommended in a descending order based on the scores of the candidate information to be recommended; and carrying out diversity sorting on the information to be recommended of the plurality of candidates sorted in front, and executing the recommendation operation corresponding to the user based on the diversity sorting result.

In the above scheme, when the information to be recommended is information published by the plurality of recalled media accounts, the recommendation module is further configured to delete the information from the same media account one by one according to ascending scores when the number of information from the same media account in the plurality of candidate information to be recommended is greater than a first threshold, until the number of information from the same media account does not exceed the first threshold; when the number of information from the same channels in the same media account number in the plurality of candidate information to be recommended is larger than a second threshold value, deleting the information from the same media account number one by one according to the ascending order of the scores until the number of information from the same media account number does not exceed the second threshold value; and when the number of the non-quality information in the candidate information to be recommended is larger than a third threshold value, deleting the non-quality information one by one according to the ascending order of the scores until the number of the non-quality information does not exceed the third threshold value.

An embodiment of the present application provides an electronic device, including:

a memory for storing executable instructions;

and the processor is used for realizing the information recommendation method based on artificial intelligence when executing the executable instructions stored in the memory.

The embodiment of the application provides a computer readable storage medium which stores executable instructions and is used for realizing the information recommendation method based on artificial intelligence.

The embodiment of the application has the following beneficial effects:

the method has the advantages that the behavior characteristics of the user are extracted from behavior data of the user corresponding to a plurality of interactive media accounts, the recalled media accounts are determined based on the behavior of the user, namely, the recalled media accounts which are in line with the user's interests are recalled through the interactive behavior of the user to the media accounts, and the method is favorable for fully mining information published by the user to the media accounts, so that information recommendation based on the recalled media accounts can be in line with the user's interests, and recommendation accuracy is effectively improved.

Drawings

FIG. 1 is a schematic diagram of an architecture of an artificial intelligence based information recommendation system 100 provided in an embodiment of the present application;

fig. 2A is a schematic structural diagram of a server 200 according to an embodiment of the present application;

Fig. 2B is a schematic structural diagram of a neural network model provided in an embodiment of the present application;

FIG. 3A is a schematic flow chart of an information recommendation method based on artificial intelligence according to an embodiment of the present application;

FIG. 3B is a schematic flow chart of an artificial intelligence based information recommendation method according to an embodiment of the present application;

FIG. 3C is a schematic flow chart of an artificial intelligence based information recommendation method according to an embodiment of the present application;

FIG. 4 is a schematic diagram of matrix decomposition provided in an embodiment of the present application;

FIG. 5 is a schematic diagram of a user behavior sequence provided in an embodiment of the present application;

FIG. 6 is a schematic diagram of an interactive media account directed graph provided by an embodiment of the present application;

FIG. 7 is a sequence diagram of interactive media accounts generated by random walks provided by embodiments of the present application;

FIG. 8 is a schematic structural diagram of a dual-tower model for implementing an artificial intelligence based information recommendation method according to an embodiment of the present application;

fig. 9 is an application scenario schematic diagram of an information recommendation method based on artificial intelligence according to an embodiment of the present application.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the present application more apparent, the present application will be described in further detail with reference to the accompanying drawings, and the described embodiments should not be construed as limiting the present application, and all other embodiments obtained by those skilled in the art without making any inventive effort are within the scope of the present application.

In the following description, reference is made to "some embodiments" which describe a subset of all possible embodiments, but it is to be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with one another without conflict.

In the following description, the terms "first", "second", "third" and the like are merely used to distinguish between similar users and do not represent a particular ordering for the users, it being understood that the "first", "second", "third" may be interchanged with a particular order or sequence, as permitted, to enable embodiments of the application described herein to be practiced otherwise than as illustrated or described herein.

In the embodiment of the application, the relevant data collection and processing should be strictly according to the requirements of relevant national laws and regulations when the example is applied, so as to obtain the informed consent or independent consent of the personal information body, and develop the subsequent data use and processing behaviors within the authorized range of the laws and regulations and the personal information body.

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of the present application only and is not intended to be limiting of the present application.

Before further describing embodiments of the present application in detail, the terms and expressions that are referred to in the embodiments of the present application are described, and are suitable for the following explanation.

1) The media account number (CP) refers to an account number for publishing articles or videos in an information stream product, some account numbers belong to official account numbers, publishing news information, etc., such as news public numbers, newspaper public numbers, some belong to self-media account numbers, and articles are published in a specific field, such as entertainment people public numbers, artificial intelligence public numbers, etc.

2) Collaborative filtering, recommending information of interest to a user by using preferences of a community of interest to each other and having common experience.

3) The graph comprises a directed graph and an undirected graph, and the graph formed by directed edges is called the directed graph.

4) The number of the edge bars of the vertex in the directed graph is called the degree of the vertex.

5) The linear rectification function (ReLU, rectified Linear Unit), also known as a modified linear unit, is a commonly used activation function (AF, activation Function) in artificial neural networks, generally referred to as a nonlinear function represented by a ramp function and its variants.

6) One-Hot encoding, also known as One-bit valid encoding, primarily uses an N-bit status register to encode N states, each of which is represented by its independent register bit and has only One bit valid at any time.

The information flow recommending products in the related technology mainly use personalized recommending technology to recommend different information according to different interests of users, wherein the information comprises articles, videos and the like. Generally, the whole recommendation flow is divided into two parts, namely recall and sorting. The aim of recall is to select partial information which is possibly interested by the user from a huge amount of information candidate pools, and combine the recalled information through various strategies to be used as input of a sequencing module. The sorting module sorts the recalled information, sorts the recalled information by using the information characteristics, the user characteristics and the cross characteristics through the output result of the scoring model, and selects articles or videos with highest scores to recommend to the user. The recall process is generally implemented by parallel processing of various strategies and models, such as searching related information based on user portraits, searching similar information based on user recent clicks, searching popular information, and the like. In the embodiment of the application, the following technical problems are found in the recall strategy related to the related technology: 1) When different types of information are presented in a mixed mode, for example, articles and videos are cross-appearing, different information can be respectively recalled by a recall algorithm in the related technology, the recall complexity is greatly increased, and interests of users on the different types of information cannot be fused. 2) In the recall strategy based on the media number, most of the recall strategy utilizes the attention behavior of the user to directly recall the media number which the user has focused on, and when the attention behavior of the user is very little, for example, the attention behavior of a new user is generally very little, the information published by the recalled media number is relatively little, and the interest requirement of the user cannot be met. 3) In the related technology, the media numbers which are popular in the current period of time are recalled, recommended to the user, all the users are covered by the popular media numbers, the recommended information deviates from the interests of the users, and poor experience is caused to the users.

In view of the above technical problems, the embodiments of the present application provide an information recommendation method, an apparatus, an electronic device, and a computer readable storage medium based on artificial intelligence, which can improve accuracy of recommendation by fusing interests of a user in various types of information by recalling a media account, and an exemplary application of the information recommendation method based on artificial intelligence provided by the embodiments of the present application is described below, where the information recommendation method based on artificial intelligence provided by the embodiments of the present application may be implemented as a server. In the following, an exemplary application when the electronic device is implemented as a server will be described.

Referring to fig. 1, fig. 1 is a schematic diagram of an architecture of an information recommendation system 100 based on artificial intelligence provided in the embodiments of the present application, where the information recommendation system may be used to support recommendation scenes of various information, such as an application scene for recommending articles, an application scene for recommending videos, an application scene for recommending media numbers of published articles and videos, etc., and according to different application scenes, the information may be articles published by media accounts, videos published by media accounts, introduction information of media accounts, etc., in the information recommendation system, a terminal 400 is connected to a server 200 through a network 300, and the network 300 may be a wide area network or a local area network, or a combination of both.

In some embodiments, the functions of the information recommendation system are implemented based on each module in the server 200, in the process that the user uses the client, the terminal 400 takes collected behavior data of the user corresponding to a plurality of interactive media account numbers as training sample data, the training sample data is collected behavior data of different users of each terminal, the neural network model is trained based on the obtained training data, and the trained neural network model is integrated in the server, wherein the neural network model comprises an extraction module 2551 and a coding module 2552; the extraction module 2551 in the server 200 extracts user behavior characteristics from behavior data of a plurality of interactive media accounts corresponding to a user, obtains user characteristic vectors through the coding module 2552, and determines a plurality of recall media accounts which satisfy similar conditions with the user characteristic vectors through the recall module 2553; the recommendation module 2554 generates information to be recommended based on the plurality of recall media accounts, performs diversified ranking processing on the information to be recommended, and performs recommendation operations of corresponding users based on the diversified ranking results.

In some embodiments, the server 200 may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDNs, and basic cloud computing services such as big data and artificial intelligence platforms. The terminal may be, but is not limited to, a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc. The terminal and the server may be directly or indirectly connected through wired or wireless communication, which is not limited in the embodiments of the present application.

Next, a structure of an electronic device for implementing an information recommendation method based on artificial intelligence provided in an embodiment of the present application is described, and as described above, the electronic device provided in the embodiment of the present application may be the server 200 in fig. 1. Referring to fig. 2A, fig. 2A is a schematic structural diagram of a server 200 provided in an embodiment of the present application, and the server 200 shown in fig. 2A includes: at least one processor 210, a memory 250, at least one network interface 220. The various components in server 200 are coupled together by bus system 240. It is understood that the bus system 240 is used to enable connected communications between these components. The bus system 240 includes a power bus, a control bus, and a status signal bus in addition to the data bus. But for clarity of illustration the various buses are labeled as bus system 240 in fig. 2A.

The processor 210 may be an integrated circuit chip with signal processing capabilities such as a general purpose processor, such as a microprocessor or any conventional processor, or the like, a digital signal processor (DSP, digital Signal Processor), or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or the like.

The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid state memory, hard drives, optical drives, and the like. Memory 250 optionally includes one or more storage devices physically located remote from processor 210.

Memory 250 includes volatile memory or nonvolatile memory, and may also include both volatile and nonvolatile memory. The non-volatile memory may be read only memory (ROM, read Only Me mory) and the volatile memory may be random access memory (RAM, random Access Memor y). The memory 250 described in embodiments of the present application is intended to comprise any suitable type of memory.

In some embodiments, memory 250 is capable of storing data to support various operations, examples of which include programs, modules and data structures, or subsets or supersets thereof, as exemplified below.

An operating system 251 including system programs for handling various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and handling hardware-based tasks; network communication module 252 for reaching other computing devices via one or more (wired or wireless) network interfaces 220, exemplary network interfaces 220 include: bluetooth, wireless compatibility authentication (WiFi), and universal serial bus (USB, universal Serial Bus), etc.

In some embodiments, the information recommendation device based on artificial intelligence provided in the embodiments of the present application may be implemented in a software manner, and fig. 2 shows the information recommendation device 255 based on artificial intelligence stored in the memory 250, which may be software in the form of a program and a plug-in, and includes the following software modules: extraction 2551, encoding 2552, recall 2553 and recommendation 2554 modules are logical and thus may be arbitrarily combined or further split depending on the functions implemented. The functions of the respective modules will be described hereinafter.

The artificial intelligence based information recommendation method provided in the embodiments of the present application will be described below in connection with exemplary applications and implementations of the server 200 provided in the embodiments of the present application. Referring to fig. 2B, fig. 2B is a schematic structural diagram of a neural network model provided in an embodiment of the present application, which may be applied to a public number recommendation system, where the neural network model may be a dual-tower model, including a user side, a media account side, and a prediction layer, where the user side and the media account side are similar in structure, and each of the user side and the media account side includes an embedding layer and a coding layer, and the coding layer includes a linear rectifying layer and a normalization layer; the prediction layer can be used for updating parameters of each layer in a training stage, and can be used for calculating the similarity between the user characteristic vector and the media account vector in an application stage so as to determine the media account vector meeting the similarity condition with the user characteristic vector and further determine the media account corresponding to the media account vector. Taking a recommendation task as an example of recommending a media account, completing the recommendation task based on each layer in the trained neural network model, namely responding to the recommendation task initiated by the terminal, acquiring behavior data of a user aiming at the media account sent by the terminal, and extracting user behavior characteristics from the behavior data through an embedded layer at the user side; rectifying the user behavior characteristics through a linear rectifying layer at the user side, and normalizing the rectifying processing result through a normalizing layer to obtain a user characteristic vector; and carrying out similarity prediction on the user feature vector and a plurality of media accounts stored offline in a media account side through a prediction layer, determining a plurality of recall media accounts meeting the similarity condition with the user feature vector, and sending the media account to be recommended to the terminal user.

In some embodiments, taking an example that the neural network model is a double-tower model, referring to fig. 8, fig. 8 is a schematic structural diagram of the double-tower model for implementing the information recommendation method based on artificial intelligence according to the embodiments of the present application. The training process of the double-tower model can be realized by the following modes: taking a combination of a user and a media account as a sample set for training a double-tower model; carrying out forward propagation on a user serving as a sample in each layer and a prediction layer in the user side of the double-tower model to obtain a user feature vector; forward spreading the media account corresponding to the user as a sample in each layer and the prediction layer in the media account side of the double-tower model to obtain a media account feature vector; determining the prediction similarity of the user feature vector and the media account feature vector; initializing a loss function comprising predicted similarity for each sample and the corresponding sample; and determining an error between the predicted similarity and the true similarity of each sample, back-propagating the error in the double-tower model according to the loss function to determine a change value of the double-tower model when the loss function obtains the minimum value, and updating parameters of the double-tower model according to the change value. The prediction layer may be implemented by a scoring function, for example, the scoring function may be cosine similarity.

The training sample of the model is in the form of a combined label of the user and the media account, namely, when the media account is the media account with the interaction behavior of the user, the true similarity of the combined sample of the user and the media account is 1, and when the media account is the media account with the interaction behavior of the user, the true similarity of the combined sample of the user and the media account is 0; the loss function may be a cross entropy loss function, a square loss function, or the like.

In some examples, the user-side embedded layer may include only identification features of the user that characterize the user behavior, and the media account-side embedded layer may include only media account identification features that characterize the user behavior.

In some examples, the user-side embedded layer may include user behavior features including at least one of identification features of a user, interactive media account features of a user, tag features of an interactive media account, channel features of an interactive media account, and the media account-side embedded layer may include media account features including at least one of identification features of a media account, tag features of a media account, channel features of a media account, and fan number features of a media account. Each user's identification feature and media account identification feature has an embedded representation of a fixed dimension when initialized. As an example, the embedded layer on the user side may include user behavior features and user attribute features.

In the following, taking a method of executing the information recommendation system provided in the embodiment of the present application by the server 200 in fig. 1 as an example, the information recommendation method based on artificial intelligence provided in the embodiment of the present application is described, and the information recommendation system includes a training phase and an application phase. First, an application of a model in the information recommendation method based on artificial intelligence provided in the embodiment of the present application is described. Referring to fig. 3A, fig. 3A is a schematic flow chart of an information recommendation method based on artificial intelligence according to an embodiment of the present application, and will be described with reference to steps 101 to 105 shown in fig. 3A.

In step 101, user behavior features are extracted from behavior data of a user corresponding to a plurality of interactive media accounts. The behavior data of the user corresponding to the plurality of interactive media accounts is clicking or watching behavior of the user aiming at the media account.

In some embodiments, extracting the user behavior feature from the behavior data of the user corresponding to the plurality of interactive media accounts may be implemented by an embedding layer in the neural network model as in fig. 2B, that is, extracting low-dimensional embedded representations from the high-dimensional original data by the embedding layer in the trained neural network model, where the low-dimensional embedded representations include at least one dimension data (e.g., a tag feature of the interactive media account, a channel feature of the interactive media account, etc.), where the embedded representation is the user behavior feature when the embedded representation is one dimension data, where the embedded representation is spliced for the plurality of dimension data when the embedded representation is the plurality of dimension data, and the spliced embedded representation is the user behavior feature.

In step 102, encoding processing is performed based on the user behavior feature to obtain a user feature vector.

In some embodiments, the encoding process based on the user behavior feature, to obtain the user feature vector, may be implemented by: rectifying the user behavior characteristics, normalizing the obtained rectifying result, and taking the obtained normalizing result as a coded user characteristic vector; wherein the user behavior feature comprises at least one of: the method comprises the steps of user interaction media account characteristics, interaction media account label characteristics and interaction media account channel characteristics.

In some examples, referring to fig. 2B, the encoding process is implemented by a user-side linear rectifying layer and a normalizing layer in the neural network model in fig. 2B. The linear rectifying layer at the user side carries out linear rectifying processing on the user behavior characteristics output by the embedded layer at the user side of the neural network model, and the linear rectifying processing can be realized by a ReLU function; then, normalization processing is performed through a normalization layer at the user side, the normalization factor is the number of neurons at the layer, and the convergence rate of the neural network model can be improved through normalization.

In some embodiments, the encoding process based on the user behavior feature, to obtain the user feature vector, may be implemented by: when the data volume of the behavior data corresponding to the plurality of interactive media accounts is smaller than a data volume threshold, or the service life of the behavior data corresponding to the plurality of interactive media accounts is not in the valid period range, extracting user attribute characteristics from attribute data of the user; and carrying out rectification treatment on the characteristics obtained after the user behavior characteristics and the user attribute characteristics are spliced and fused, carrying out normalization treatment on the obtained rectification result, and taking the obtained normalization treatment result as a user characteristic vector obtained by encoding.

In the embodiment of the application, when the behavior data of the user corresponding to a plurality of interactive media accounts is insufficient or exceeds the period, the attribute characteristics are supplemented, so that the defect that the learned behavior characteristics of the user are insufficient based on the behavior data of which the data quantity is insufficient or exceeds the period is overcome, and the learning precision of the neural network model is improved; and carrying out coding processing on the characteristics obtained after the user behavior characteristics and the user attribute characteristics are spliced and fused, so that the user characteristic vector obtained by coding can be better mapped to the interests and the demands of the user.

In step 103, a plurality of recall media accounts satisfying a similar condition to the user feature vector is determined.

In some embodiments, referring to fig. 3B, fig. 3B is a schematic flow chart of an information recommendation method based on artificial intelligence according to an embodiment of the present application, which illustrates step 103 in fig. 3A, and may also be implemented by performing steps 1031 to 1033. The steps will be described.

In step 1031, media account feature vectors for a plurality of candidate media accounts are obtained.

In some examples, obtaining media account feature vectors for a plurality of candidate media accounts may be accomplished by: for each of the plurality of candidate media accounts, performing the following: extracting media account characteristics from attribute data of candidate media accounts; rectifying the characteristics of the media account, normalizing the obtained rectifying result, and taking the obtained normalizing result as a characteristic vector of the media account; wherein the media account characteristics include at least one of: the label characteristics of the media account, the channel characteristics of the media account and the vermicelli quantity characteristics of the media account.

In some examples, referring to fig. 2B, the processing performed for each of the plurality of candidate media accounts is implemented by an embedding layer, a linear rectifying layer, and a normalizing layer on the media account side in the neural network model in fig. 2B. The embedding layer of the media account side is used for extracting media account characteristics from attribute data of candidate media accounts; the linear rectifying layer at the media account side carries out linear rectifying treatment on the characteristics of the media account output through the embedded layer, and the linear rectifying treatment can be realized through a ReLU function; then, normalization processing is carried out through a normalization layer at the media account side, the normalization factor is the number of neurons of the layer, and the convergence rate of the double-tower model can be improved through normalization.

It should be noted that, the candidate media account may only include a media account having interaction with the user (i.e., an interactive media account), and as an example, in a training stage of the neural network model, that is, when the interactive media account is subjected to forward operation, a media account vector of the interactive media account is stored offline and stored in a vector database. And in the online application stage, when the recalled media account meeting the similarity condition with the user feature vector is determined, a plurality of media accounts with the nearest similarity are directly retrieved by using a vector database, and related recommendation is carried out.

In the embodiment of the present application, when the candidate media account includes only the interactive media account, the media account may be stored offline, so as to quickly find the media account with the same interest as the user based on the user feature vector.

In some embodiments, the recalled candidate media accounts may include media accounts that have an interaction with the user (i.e., interactive media accounts) and media accounts that have no interaction with the user (i.e., non-interactive media accounts); the recalled candidate media accounts may also include only media accounts for which there is no interaction with the user.

Taking the example that the candidate media account only comprises a non-interactive media account, the behavior characteristics of the user are extracted from the behavior data aiming at the interactive media account 1-10, and in the recall stage, recall is carried out from the media account 11-100, so that a plurality of recall media accounts are obtained. Taking the example that the candidate media accounts comprise an interactive media account and a non-interactive media account, the behavior characteristics of the user are extracted from the behavior data aiming at the interactive media accounts 1-10, and in the recall stage, recall is carried out from the media accounts 1-100, so that a plurality of recall media accounts are obtained.

In the embodiment of the invention, when the candidate media account only comprises the non-interactive media account, the candidate media account is multiplexed into the media account which does not have interaction with the user by learning the characteristics from the interactive media account, so that the media account which has similar interests with the user and is fresh can be effectively mined.

In step 1032, cosine distances between the media account feature vectors and the user feature vectors of the plurality of candidate media accounts are determined, with the cosine distances being used as the similarity.

It should be noted that, in some examples, the similarity may also be calculated by pearson correlation coefficient, mahalanobis distance, euclidean distance, or the like.

In step 1033, a plurality of candidate media accounts having a similarity to the user feature vector exceeding a similarity threshold are determined as a plurality of recalled media accounts satisfying the similarity condition.

In some examples, multiple recall media accounts satisfying similar conditions may also be determined by: and determining the similarity between the candidate media account vectors and the user feature vectors, and using the candidate media accounts corresponding to the candidate media account vectors with the similarity ranked at the front as a plurality of recall media accounts meeting the similarity condition.

For example, the top candidate media account number vector may be obtained by a top number or proportion, e.g., the top 50 candidate media account number vectors may be obtained, or the top two percent of the total number of all media account number vectors may be obtained. In step 104, information to be recommended is generated based on the plurality of recall media accounts.

In some embodiments, the information to be recommended may be generated by at least one of: generating information to be recommended for recommending the media account based on the plurality of recall media accounts; based on the information published by the plurality of recall media accounts, information to be recommended for recommending the information published by the media accounts is generated.

In some examples, based on the multiple recall media accounts, generating the information to be recommended for recommending the media account is to directly use the recalled media account as the information to be recommended, such as public numbers of articles and videos. Based on the information published by the plurality of recall media accounts, generating information to be recommended for recommending the information published by the media accounts, namely carrying out diversity sorting processing on the information published by the plurality of recall media accounts, and taking the information published by the recall media accounts after the diversity sorting processing as the information to be recommended, such as articles, videos and the like.

In step 105, a recommendation operation of the corresponding user is performed based on the information to be recommended.

In some embodiments, referring to fig. 3C, fig. 3C is a schematic flow chart of an artificial intelligence based information recommendation method provided in an embodiment of the present application, which illustrates step 105 in fig. 3A, and may also be implemented by performing steps 1051 to 1053. The steps will be described.

In step 1051, a scoring model is invoked to predict a score of the information to be recommended based on the information feature vector of the information to be recommended, the user feature vector of the user, and the cross feature vector of the user and the information to be recommended, the score characterizing the similarity of the information to be recommended and the user.

In some examples, a scoring model used in the information recommendation method based on artificial intelligence and training performed by the scoring model provided by the embodiment of the application are described, wherein the scoring model comprises a feature extraction module, a fusion coding module and a prediction module. The training process of the scoring model can be realized by the following modes: taking the combination of the user and the information to be recommended as a sample set for training a double-tower model, and initializing a loss function comprising each sample and a prediction score of the corresponding sample; extracting information features, user features of the user and cross features of the user and information to be recommended from the combined sample through a feature extraction module; the information feature of the information to be recommended, the user feature of the user and the cross feature of the information to be recommended are coded through a fusion coding module to obtain an information feature vector of the information to be recommended, a user feature vector of the user and the cross feature vector of the information to be recommended, and the information feature vector of the information to be recommended, the user feature vector of the user and the cross feature vector of the information to be recommended are fused, for example, the information feature vector of the information to be recommended, the user feature vector of the user and the cross feature vector of the information to be recommended are connected through full connection; and predicting the connection result through a prediction module to obtain a score of the information to be recommended, determining an error between the prediction score and the true score of each sample, reversely propagating the error in the scoring model according to the loss function to determine a change value of the scoring model when the loss function obtains the minimum value, and updating parameters of the scoring model according to the change value.

It should be noted that, the prediction model may be implemented by a logistic regression function softmax or a cosine similarity function, and when implemented by the cosine similarity function, the score may represent the similarity between the information to be recommended and the user. The cross characteristic of the user and the information to be recommended is obtained by simply string splicing the original data of the user and the information to be recommended and performing One-Hot coding on the spliced result.

In the embodiment of the application, the scoring model is realized through a neural network model. In the neural network model, the training sample is in the form of a combined label of the user and the information to be recommended, namely, when the information to be recommended is information clicked or watched by the user, the real similarity of the combined sample of the user and the recommended information is 1, and when the information to be recommended is information not clicked and watched by the user, the real similarity of the combined sample of the user and the recommended information is 0; the loss function may be a cross entropy loss function, a square loss function, or the like.

In step 1052, the candidate information to be recommended is sorted in descending order based on the scores of the candidate information to be recommended.

In step 1053, the information to be recommended of the plurality of candidates ranked in front is subjected to diversity ranking, and the recommendation operation of the corresponding user is performed based on the diversity ranking result. Here, the recommendation operation for the user is: and sending a plurality of pieces of information to be recommended which are ordered according to the ordering result to the user terminal.

In some examples, when the information to be recommended is information published by multiple recall media accounts, the diversity ranking of the top ranked multiple candidate users may be achieved by at least one of: when the number of information from the same media account number in the plurality of candidate information to be recommended is larger than a first threshold value, deleting the information from the same media account number one by one according to the ascending order of the scores until the number of information from the same media account number does not exceed the first threshold value; when the number of information from the same channels in the same media account number in the plurality of candidate information to be recommended is larger than a second threshold value, deleting the information from the same media account number one by one according to the ascending order of the scores until the number of information from the same media account number does not exceed the second threshold value; and when the number of the non-quality information in the candidate information to be recommended is larger than a third threshold value, deleting the non-quality information one by one according to the ascending order of the scores until the number of the non-quality information does not exceed the third threshold value. The first threshold, the second threshold, and the third threshold may be determined according to the total amount of information to be recommended that is recalled.

Taking information to be recommended as an article example, when the number of articles from the same public number in the plurality of candidate articles is greater than 50, deleting the articles from the same public number one by one in ascending order of the article scores until the number of articles from the same public number is not more than 50, and ensuring that the number of articles from the same public number in the plurality of candidate articles is at most 50.

In the embodiment of the application, the information to be recommended is subjected to diversity sorting, so that the diversity of the information to be recommended is ensured, the problem that the aggregation effect occurs on the head of the information to be recommended is solved, the information to be recommended is ensured to be scattered in different media accounts and channels and labels of different media accounts, namely, the recalled information to be recommended is ensured to be sourced from a plurality of media accounts and various types of media accounts, and the balanced exposure opportunity of the media accounts is given; and meanwhile, the quality of the information to be recommended is also ensured.

Since user behavior typically occurs between a user and media information (e.g., articles, videos, etc.), embodiments of the present application map behavior data of the user for the media information to behavior data of the user for the media account, for example, user u clicks on article d, which is published by media account c, and then user u has interaction with media account c. In some embodiments, before extracting the behavior characteristics of the user from the behavior data of the user corresponding to the plurality of interactive media accounts, the behavior data of the user corresponding to the plurality of interactive media accounts may be further obtained, that is, the behavior data of the user corresponding to the plurality of information is mapped to the behavior data of the user corresponding to the interactive media account for publishing the information; wherein the behavior data of the user for the plurality of information characterizes at least one of the following behaviors: and the user focuses on the behavior of the information published by the interactive media account and the behavior of subscribing the information published by the interactive media account.

In some embodiments, when the obtained data size of the behavior data of the user corresponding to the plurality of interactive media account numbers is insufficient, the user behavior feature obtained based on the behavior data cannot accurately represent the behavior of the user, so the following processing may be performed before the encoding processing is performed based on the user behavior feature to obtain the user feature vector, so as to obtain the user behavior feature capable of accurately reflecting the user behavior: determining user collaborative filtering characteristics and user graph characteristics, sequentially connecting the user collaborative filtering characteristics, the user graph characteristics and the user behavior characteristics obtained through extraction, and taking the new user behavior characteristics obtained through connection as user behavior characteristics for coding processing; wherein the user collaborative filtering feature is determined by decomposing a matrix comprising behavior data, and the user graph feature is determined based on mapping a directed graph comprising behavior data.

In some examples, collaborative filtering characteristics may be determined by: based on behavior data of users aiming at a plurality of interactive media accounts, constructing an interaction matrix corresponding to each interactive media account of each user; and decomposing the interaction matrix to obtain user collaborative filtering characteristics representing the user behaviors and media account collaborative filtering characteristics representing the user behaviors.

For example, referring to fig. 4, fig. 4 is a schematic diagram of matrix decomposition provided in an embodiment of the present application. Where 401 is the interaction matrix, 402 is the user collaborative filtering feature, and 403 is the media account collaborative filtering feature. The interaction matrix 401 of the user and the media account is a matrix 3*5, which represents that the number of users is 3, the number of media accounts is 5, and the high-dimensional interaction matrix is decomposed into two low-dimensional matrices through matrix decomposition, namely, the matrix of the user matrix 402 is 3*6, which can be understood as vector representations of 3 users, namely, the vector representations of the 3 users are user collaborative filtering characteristics, the vector representation of the media account matrix 403 is a matrix 6*5, which can be understood as vector representations of 5 media accounts, namely, the vector representations of the 5 media accounts are media account collaborative filtering characteristics.

Because the scale of the user and the media account is very large in the actual scene, in the embodiment of the application, the high-dimensional interaction matrix is subjected to matrix decomposition, so that dense user collaborative filtering characteristics and media account collaborative filtering characteristics can be obtained, the problem of sparse data is well solved, and the user and the media account are in the same vector space, so that recall processing is performed on the user behavior characteristics which are determined based on the user collaborative filtering characteristics and the media account collaborative filtering characteristics, and the recall precision is higher.

In some examples, the user graph features may be determined by: acquiring a behavior sequence of each user in a time window; the behavior sequence characterizes the sequence of interactive behaviors of the user with the media account in a time window; and constructing a directed graph of the interactive media accounts of each user by taking the media account with interactive behaviors of the user as the top points and turning the media accounts related in the behavior sequence as the edges between the top points, generating a sequence of the interactive media accounts based on the directed graph, obtaining an embedded representation of each interactive media account based on the sequence of each interactive media account, and taking the embedded representation as the characteristics of the user graph.

For example, referring to fig. 5, fig. 5 is a schematic diagram of a user behavior sequence provided in an embodiment of the present application. Since the interests of a user may change over time, the user's behavior within a set time window is taken as a sequence of user behaviors. In fig. 5, a sequence of three users' behaviors over a certain time window is shown, with 1-5 representing different media accounts. A directed graph is then constructed for the sequence of user actions. Referring to fig. 6, fig. 6 is a schematic diagram of an interactive media account directed graph provided in an embodiment of the present application. Two adjacent interactive media accounts within a time window may be connected by a directed edge, for example, in fig. 3, the user a sequentially accesses the interactive media account 1 and the interactive media account 2, that is, the interactive media account 1 and the interactive media account 2 have a directed edge, where the directed edge is the direction of the interactive media account 1 to the interactive media account 2. And distributing corresponding weights to the directed edges in the directed graph through the cooperative behaviors of all users. The weight of the directed edges of the interactive media account 1 to the interactive media account 2 is equal to the proportion of the frequency of the interactive media account 1 turning to the interactive media account 2 to the outgoing degree of the interactive media account 1. Then, a sequence of a plurality of interactive media accounts is generated by using a random walk concept according to the constructed directed graph of the interactive media account, referring to fig. 7, fig. 7 is a schematic diagram of the sequence of the interactive media account generated by the random walk according to the embodiment of the present application, where 701 shows the sequence of the interactive media account 1, 702 is the sequence of the interactive media account 2, and 703 is the sequence of the interactive media account 3. An embedded representation of each interactive media account associated with the user may be derived based on the sequence of each interactive media account, and the embedded representation may be characterized as a user graph.

In the embodiment of the application, the graph characteristics combined with the time sequence information are obtained by constructing the directed graph, so that the problem of data sparseness is well solved.

In some embodiments, after extracting the media account features from the attribute data of the candidate media account, the following processing may also be performed to obtain dense media account features: determining media account collaborative filtering characteristics and media account map characteristics of candidate media accounts, connecting the media account collaborative filtering characteristics, the media account map characteristics and the media account characteristics obtained through extraction, and taking the new media account characteristics obtained through connection as the media account characteristics for rectifying treatment; wherein the media account collaborative filtering characteristics are determined by decomposing a matrix comprising behavior data, and the media account graph characteristics are determined based on mapping a directed graph comprising behavior data.

It should be noted that the collaborative filtering feature of the media account is determined by decomposing the interaction matrix as described above; the manner in which the media account map features are determined is the same as the manner in which the user account map features are determined.

In the following, an exemplary application of the embodiments of the present application in a practical application scenario will be described.

Referring to fig. 9, fig. 9 is an application scenario schematic diagram of an information recommendation method based on artificial intelligence according to an embodiment of the present application. A specific implementation scenario of the information recommendation method based on artificial intelligence according to the embodiment of the present application will be described below with reference to fig. 9.

In this embodiment of the present application, taking an information flow product as an example, the information recommendation system maps clicking or viewing behaviors of users on different types of information such as articles and videos to interaction behaviors of users on media numbers (i.e. media account numbers of articles, videos, etc., hereinafter abbreviated as CPs), so as to recall the media numbers in combination with user interests. That is, the information recommendation system maps the user behavior between the user and the article and the user behavior between the user and the video in the information stream product to the user and the CP, for example, when the user u clicks the article d, d is issued by the media number c, then the interaction behavior between u and c exists.

In some embodiments, recall of media numbers by the information recommendation system may be implemented through a dual tower model. Referring to fig. 9, the dual-tower model encodes the user-side feature and the CP-side feature, respectively, and fits the similarity between the encoded results of the user-side feature and the CP-side feature by a scoring function, which may be a cosine function. After the training of the double-tower model is completed, performing forward calculation on all CPs offline to obtain codes of the CPs and storing the codes; in online calculation, forward calculation is carried out on a user to obtain the code of the user, then the most similar Top N CPs are retrieved from the CPs stored offline, and the Top N CPs are used as recalled media numbers.

The structure of the double-tower model comprises an embedding layer, a coding layer and a prediction layer, wherein the embedding layer is firstly used for embedding the embedding layer, the embedding representations of the user-emb, namely the user feature user emb embedding are spliced, for example, in FIG. 9, the embedding representation of the CP-emb, namely the media number feature CP embdding, the embedding representation of the tag-emb, namely the tag feature tag embedding of the media number, the embedding representation of the chanl-emb, namely the media number feature chan nel embedding are spliced, and the embedding representations of the multiple dimensions are spliced to obtain the user feature, the media number feature and the media account feature; and then the encoding layer captures the characteristics of the embedding layer by using a full-connection network, the encoding layer comprises a linear rectifying layer and a normalizing layer, the linear rectifying layer provides nonlinear change after full connection by using Re LU (linear rectifying function) as an activation function, gradient descent and counter propagation are facilitated, and the normalizing layer normalizes the linear rectifying result. It should be noted that, in some embodiments, a dual-tower network may be constructed by using only the embedded representation of the user identification feature user-id and the media number identification feature CP-id, but since many users or media numbers have fewer interactions, the corresponding ebedding cannot be accurately learned, and thus, the information recommendation system adds various side information to enhance the ebedding process. Specific features include: 1) CP-side features: CP-id, i.e. the identifying feature of CPs, each CP has an ebedding representation of a fixed dimension, which can be obtained by looking up a table during training or prediction; tag-id, namely a tag feature in a CP image, wherein one CP generally comprises a plurality of tags, and when a plurality of tag features of an interactive media number of a user exist, the tag features are generally taken as tag features of the interactive media number on average; chanl-id, namely channel characteristics in the CP portrait; fan (not shown in the figure), namely the vermicelli number feature of CP, since the vermicelli number is a continuous feature, the vermicelli number can be segmented, and each number is divided into one segment, so as to convert the continuous vermicelli number feature into a discrete vermicelli number feature. 2) user side features: the user-id characteristic, namely the identification characteristic of the user, can be obtained through a table look-up as the identification characteristic of the CP; the CP-id is the interactive media number feature of the user, the CP which has the interactive behavior (clicking or watching) recently with the user is taken, and when a plurality of interactive media number features of the user exist, the average is taken as the interactive media number feature of the user; tag-id, namely the tag characteristic of the interactive media number, counts the tags of the CP which have the latest interactive behavior of the user, takes the average as the tag characteristic of the interactive media number, and can carry out average processing on the tags of the interactive media numbers which are ranked at the front of the times of the interactive behavior in the time period; the chanl-id, namely the channel characteristics of the interactive media number, is counted, the average is taken as the channel characteristics of the interactive media number, and the average can be used for carrying out average processing on the channels of the interactive media numbers with the top ranking times of the interactive behaviors in the time period.

The portraits of the media numbers mainly include tags, channels, and the like, and are extracted from media number publication articles or videos. For example, general articles have some static features, channels, labels, etc., called representations of the articles. And the information recommendation system counts the static characteristics of the articles corresponding to the CPs to obtain the image information of the CPs. However, the article image is distinguished from the media number image, and taking the channel feature as an example, one article generally has only one channel, but the text type of the media number is likely to be not fixed, so that a plurality of channels are possible, and after statistics, a plurality of head parts are taken as the channels in the media number image.

In some embodiments, besides using random initialization before training, the embedding layer processing in the information recommendation system introduces more global embedding information to make up for the defect of sparse user or insufficient CP learning. Mainly comprises two kinds of collaborative filtering characteristics cf-emmbedding and graph-emmbedding. Wherein, cf-casting: the emmbedding of the behavior-similar CP should be similar, as should the emmbedding of the behavior-similar user. The related art user-based nearest neighbor UCF recommendation and the object-based ICF recommendation actually construct a user vector space and a media number vector space respectively, and can be recommended by finding similarity in any vector space. According to the embodiment of the application, the high-dimensional matrix can be mapped into the product of two low-dimensional matrices through matrix decomposition, so that dense user and CP vectors can be obtained, namely, the user collaborative filtering characteristics and the media number collaborative filtering characteristics are obtained, the problem of data sparseness is well solved, the user and the CP are in the same vector space, and the prediction accuracy is good. graph-embedding: reflecting the sequential information of the user's interactive media number (which may more precisely influence the user's preference), the information recommendation system regards the user's behavior within a set time window as a sequence of user behaviors, since the interests of one user may drift over time. The sequence of the plurality of interactive media numbers is obtained as above, and the em bearing representation, i.e. the graph feature, of each CP is learned based on the sequence of each interactive media account. In the embodiment of the application, the cf-EMbe and graph-EMbe learned in the two modes are used in a splicing mode, and are not involved in updating in the process of training the double-tower model.

In some embodiments, different embading and double-tower models are trained in an offline training process, when the double-tower model is trained, forward computation is performed on all CPs to obtain media number feature vectors, a vector database fasss index is constructed based on the media number feature vectors, and the media number feature vectors and corresponding media numbers are stored offline. In the online calculation process, forward calculation is performed on the user in real time to obtain a user feature vector, and the nearest top N (top N ranking) CPs are obtained by utilizing fass retrieval. The information recommendation system can directly recommend the obtained CP to the user, and also can recommend articles or videos corresponding to the CP to the user.

In some embodiments, a CP may typically post a lot of articles or videos, for example, when the information recommendation system recalls a lot of CPs, the candidate articles are too many, and the most suitable articles are selected by the scoring model. The features of the scoring model may use article features, user-to-article cross features (e.g., media number features of a publication article). As an example, specific features used may include: 1) CP number of vermicelli and high quality article ratio; 2) The text time, the quality score of the article and the new heat of the article; 3) Intersection ratio of article portraits and user portraits.

Here, the article portrait includes tag features, channel features, etc. of media numbers of the articles, and the user portrait includes user behavior data and user attribute data, for example, tag features, channel features, etc. of interactive media account numbers in the behavior data. The weights of the features can construct training data for training and learning, and can be manually specified according to actual business, for example, an information recommendation system can improve weights of three features, namely recalled high-quality articles, CP high-quality articles duty ratio and articles quality score, so that the recalled articles are strongly related to user interests.

In some embodiments, the information recommendation system can obtain a ranked article list through recall of the CP and scoring of the articles, and can also add diversity processing to adjust the ranking, so as to ensure diversity of recall results, and solve the problem that the head of the articles has aggregation effect caused by the scoring model, namely, the recalled articles are all derived from one or a plurality of media numbers, so that a plurality of media numbers have no exposure opportunity. The diversity sorting process specifically comprises the following steps: 1) Diversity of CP: ensuring that the articles corresponding to each media number do not exceed a certain threshold, wherein the threshold is determined according to the actual business recall number; 2) Diversity of channels: ensuring that the recall articles are scattered in different channels, wherein the article quantity of each channel can be obtained by multiplying the total quantity of recall sections by the duty ratio of each channel in the user portrait; 3) And (3) quality control: in order to ensure exposure of quality media numbers, a ranking module in the information recommendation system limits the number of non-quality media numbers.

In the embodiment of the application, the behaviors of the users on the articles and the videos are mapped to the interests of the users on the CPs, the interests of various types of information are comprehensively utilized, and the interests of the users are mined more fully; covering various application scenes, the method can support recommendation of media numbers, recommendation of articles and videos, and simultaneous recommendation of articles and videos; by utilizing various side information and global characteristic information, additional ebedding representations of users and media numbers are learned, so that the problems of sparse behavior, cold start and the like of a recommended scene can be effectively relieved, and the accuracy of recommendation is enhanced.

Continuing with the description below of an exemplary architecture provided by embodiments of the present application for implementing the artificial intelligence based information recommendation device 255 as a software module, in some embodiments, as shown in FIG. 2A, the software module stored in the artificial intelligence based information recommendation device 255 of the memory 250 may include:

an extracting module 2551, configured to extract a user behavior feature from behavior data of a plurality of interactive media accounts corresponding to a user; the encoding module 2552 is configured to perform encoding processing based on the user behavior feature to obtain a user feature vector; a recall module 2553 configured to determine a plurality of recall media accounts that satisfy a similar condition to the user feature vector; and the recommending module 2554 is configured to generate information to be recommended based on the multiple recall media account numbers, and execute a recommending operation corresponding to the user based on the information to be recommended.

In some embodiments, the extracting module 2551 is further configured to map behavior data of the user for a plurality of information into behavior data of the user for an interactive media account on which the information is published; wherein the behavior data of the user for a plurality of information characterizes at least one of the following behaviors: and the user focuses on the behavior of the information published by the interactive media account and the behavior of subscribing the information published by the interactive media account.

In some embodiments, the encoding module 2552 is further configured to determine a user collaborative filtering feature and a user graph feature, sequentially connect the user collaborative filtering feature, the user graph feature, and the user behavior feature obtained by extraction, and use the new user behavior feature obtained by connection as a user behavior feature for performing the encoding process; wherein the user collaborative filtering feature is determined by decomposing a matrix comprising the behavior data, and the user graph feature is determined based on mapping a directed graph comprising the behavior data.

In some embodiments, the encoding module 2552 is further configured to extract user attribute features from the user's attribute data; rectifying the user behavior characteristics and the user attribute characteristics, or rectifying the user behavior characteristics, normalizing the obtained rectifying result, and taking the obtained normalizing result as a user characteristic vector obtained by encoding; wherein the user behavior feature comprises at least one of: the method comprises the steps of enabling the user to interact with a media account number, enabling the user to interact with the media account number, and enabling the user to interact with the media account number.

In some embodiments, the recall module 2553 is further configured to obtain media account feature vectors of a plurality of candidate media accounts, determine cosine distances between the media account feature vectors of the plurality of candidate media accounts and the user feature vector, and use the cosine distances as the similarity; and determining a plurality of candidate media accounts with similarity exceeding a similarity threshold value with the user feature vector as a plurality of recall media accounts meeting the similarity condition.

In some embodiments, the recall module 2553 is further configured to, for each of the plurality of candidate media accounts, perform the following: extracting media account characteristics from attribute data of the candidate media account; rectifying the media account characteristics, normalizing the obtained rectifying result, and taking the obtained normalizing result as a media account characteristic vector; wherein the media account characteristics include at least one of: the label characteristics of the media account, the channel characteristics of the media account and the vermicelli quantity characteristics of the media account.

In some embodiments, the recall module 2553 is further configured to determine a media account collaborative filtering feature and a media account map feature of the candidate media account, connect the media account collaborative filtering feature, the media account map feature, and the media account feature obtained by extraction, and use the new media account feature obtained by connection as a media account feature for performing the rectification processing; wherein the media account collaborative filtering feature is determined by decomposing a matrix comprising the behavior data, and the media account graph feature is determined based on mapping a directed graph comprising the behavior data.

In some embodiments, the recommending module 2554 is further configured to generate the information to be recommended by at least one of: generating information to be recommended for recommending the media account based on the plurality of recall media accounts; generating information to be recommended for recommending the information published by the media account based on the information published by the plurality of recall media accounts; predicting to obtain a score of the information to be recommended based on the information feature vector of the information to be recommended, the user feature vector of the user and the cross feature vector of the user and the information to be recommended, wherein the score characterizes the similarity of the information to be recommended and the user; sorting the candidate information to be recommended in a descending order based on the scores of the candidate information to be recommended; and carrying out diversity sorting on the information to be recommended of the plurality of candidates sorted in front, and executing the recommendation operation corresponding to the user based on the diversity sorting result.

In some embodiments, when the information to be recommended is information published by the plurality of recalled media accounts, the recommendation module 2554 is further configured to, when the number of information from the same media account in the plurality of candidate information to be recommended is greater than a first threshold, delete the information from the same media account one by one in ascending order of scores until the number of information from the same media account does not exceed the first threshold; when the number of information from the same channels in the same media account number in the plurality of candidate information to be recommended is larger than a second threshold value, deleting the information from the same media account number one by one according to the ascending order of the scores until the number of information from the same media account number does not exceed the second threshold value; and when the number of the non-quality information in the candidate information to be recommended is larger than a third threshold value, deleting the non-quality information one by one according to the ascending order of the scores until the number of the non-quality information does not exceed the third threshold value.

Embodiments of the present application provide a computer program product or computer program comprising 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 executes the information recommendation method based on artificial intelligence according to the embodiment of the application.

The embodiments of the present application provide a computer readable storage medium having stored therein executable instructions that, when executed by a processor, cause the processor to perform the artificial intelligence based information recommendation method provided by the embodiments of the present application, for example, as shown in fig. 3A, 3B, 3C.

In some embodiments, the computer readable storage medium may be FRAM, ROM, PROM, EP ROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; but may be a variety of devices including one or any combination of the above memories.

In some embodiments, the executable instructions may be in the form of programs, software modules, scripts, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and they may be deployed in any form, including as stand-alone programs or as modules, components, subroutines, or other units suitable for use in a computing environment.

As an example, the executable instructions may, but need not, correspond to files in a file system, may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a hypertext markup language (HTML, hyper Text Markup Language) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub-programs, or portions of code).

As an example, executable instructions may be deployed to be executed on one computing device or on multiple computing devices located at one site or, alternatively, distributed across multiple sites and interconnected by a communication network.

In summary, according to the embodiment of the application, when the behavior data of the user corresponding to the plurality of interactive media accounts is insufficient or exceeds the period, the attribute characteristics are supplemented, so that the defect that the learned behavior characteristics of the user are insufficient based on the behavior data of which the data quantity is insufficient or exceeds the period is overcome, and the learning precision of the neural network model is improved; the characteristics obtained after the user behavior characteristics and the user attribute characteristics are spliced and fused are subjected to coding processing, so that the user characteristic vector obtained by coding can be better mapped to interests and demands of a user; the high-dimensional interaction matrix is subjected to matrix decomposition, dense user collaborative filtering characteristics and media account collaborative filtering characteristics can be obtained, the problem of sparse data is well solved, and the user and the media account are in the same vector space, so that recall processing is carried out on the user behavior characteristics determined based on the user collaborative filtering characteristics and the media account collaborative filtering characteristics, and the recall precision is higher; by means of constructing the directed graph, graph characteristics combined with time sequence information are obtained, and the problem of data sparseness is well solved.

The foregoing is merely exemplary embodiments of the present application and is not intended to limit the scope of the present application. Any modifications, equivalent substitutions, improvements, etc. that are within the spirit and scope of the present application are intended to be included within the scope of the present application.

Claims

1. An artificial intelligence based information recommendation method, comprising:

extracting user behavior characteristics from behavior data of a plurality of interactive media accounts corresponding to the user;

coding processing is carried out based on the user behavior characteristics to obtain user characteristic vectors;

determining a plurality of recall media accounts satisfying similar conditions with the user feature vector;

generating information to be recommended based on the plurality of recall media accounts and at least one of information published by the plurality of recall media accounts;

the information to be recommended of a plurality of candidates is subjected to diversity sorting, wherein the diversity sorting is performed according to at least one of the following parameters of the information to be recommended of the plurality of candidates: the number of information from the same media account, the number of information from the same channel in the same media account, the number of non-premium information;

and executing the recommendation operation corresponding to the user based on the result of the diversity sorting.

2. The method of claim 1, wherein prior to extracting user behavioral characteristics from the behavioral data of the user corresponding to the plurality of interactive media account numbers, the method further comprises:

mapping behavior data of the user aiming at a plurality of pieces of information into behavior data of the user aiming at an interactive media account for publishing the information;

wherein the behavior data of the user for a plurality of information characterizes at least one of the following behaviors: and the user focuses on the behavior of the information published by the interactive media account and the behavior of subscribing the information published by the interactive media account.

3. The method of claim 1, wherein prior to performing the encoding process based on the user behavior feature to obtain a user feature vector, the method further comprises:

determining user collaborative filtering characteristics and user graph characteristics, sequentially connecting the user collaborative filtering characteristics, the user graph characteristics and the user behavior characteristics obtained through extraction, and taking the new user behavior characteristics obtained through connection as user behavior characteristics for coding processing;

wherein the user collaborative filtering feature is determined by decomposing a matrix comprising the behavior data, and the user graph feature is determined based on mapping a directed graph comprising the behavior data.

4. The method of claim 1, wherein the encoding based on the user behavior feature to obtain a user feature vector comprises:

extracting user attribute characteristics from the attribute data of the user;

rectifying the user behavior characteristics and the user attribute characteristics, or rectifying the user behavior characteristics, normalizing the obtained rectifying result, and taking the obtained normalizing result as a user characteristic vector obtained by encoding;

wherein the user behavior feature comprises at least one of: the method comprises the steps of enabling the user to interact with a media account number, enabling the user to interact with the media account number, and enabling the user to interact with the media account number.

5. The method of claim 1, wherein the determining a plurality of recall media accounts that satisfy a similar condition to the user feature vector comprises:

acquiring media account feature vectors of a plurality of candidate media accounts, determining cosine distances between the media account feature vectors of the plurality of candidate media accounts and the user feature vectors, and taking the cosine distances as similarity;

and determining a plurality of candidate media accounts with similarity exceeding a similarity threshold value with the user feature vector as a plurality of recall media accounts meeting the similarity condition.

6. The method of claim 5, wherein the obtaining media account feature vectors for a plurality of candidate media accounts comprises:

for each candidate media account of the plurality of candidate media accounts, performing the following:

extracting media account characteristics from attribute data of the candidate media account;

rectifying the media account characteristics, normalizing the obtained rectifying result, and taking the obtained normalizing result as a media account characteristic vector;

wherein the media account characteristics include at least one of: the label characteristics of the media account, the channel characteristics of the media account and the vermicelli quantity characteristics of the media account.

7. The method of claim 6, wherein after extracting media account features from the attribute data of the candidate media account, the method further comprises:

determining media account collaborative filtering characteristics and media account map characteristics of the candidate media accounts, connecting the media account collaborative filtering characteristics, the media account map characteristics and the media account characteristics obtained through extraction, and taking the new media account characteristics obtained through connection as the media account characteristics for carrying out rectification processing;

Wherein the media account collaborative filtering feature is determined by decomposing a matrix comprising the behavior data, and the media account graph feature is determined based on mapping a directed graph comprising the behavior data.

8. The method of claim 1, wherein the generating information to be recommended based on the plurality of recall media accounts and at least one of the information published by the plurality of recall media accounts comprises:

generating information to be recommended by at least one of the following means:

generating information to be recommended for recommending the media account based on the plurality of recall media accounts;

generating information to be recommended for recommending the information published by the media account based on the information published by the plurality of recall media accounts;

the diversity sorting of the information to be recommended of the plurality of candidates includes:

predicting to obtain a score of the information to be recommended based on the information feature vector of the information to be recommended, the user feature vector of the user and the cross feature vector of the user and the information to be recommended, wherein the score characterizes the similarity of the information to be recommended and the user;

sorting the information to be recommended of the plurality of candidates in a descending order based on scores of the information to be recommended of the plurality of candidates;

And carrying out diversity sorting on the information to be recommended of the plurality of candidates sorted in the front.

9. The method of claim 8, wherein the diversity ranking of the information to be recommended for the top ranked plurality of candidates comprises at least one of:

when the number of information from the same media account in the plurality of candidate information to be recommended is larger than a first threshold, deleting the information from the same media account one by one according to the ascending order of the score until the number of information from the same media account does not exceed the first threshold;

when the number of the information from the same channels in the same media account number in the plurality of candidate information to be recommended is larger than a second threshold value, deleting the information from the same media account number one by one according to the ascending order of the scores until the number of the information from the same media account number does not exceed the second threshold value;

and when the number of the non-quality information in the information to be recommended of the plurality of candidates is larger than a third threshold value, deleting the non-quality information one by one according to the ascending order of the scores until the number of the non-quality information does not exceed the third threshold value.

10. An artificial intelligence based information recommendation device, comprising:

the extraction module is used for extracting user behavior characteristics from behavior data of the user corresponding to the plurality of interactive media account numbers;

the coding module is used for carrying out coding processing based on the user behavior characteristics to obtain user characteristic vectors;

a recall module for determining a plurality of recall media accounts satisfying a similar condition to the user feature vector;

the recommendation module is used for generating information to be recommended based on the plurality of recall media accounts and at least one of information published by the plurality of recall media accounts; the information to be recommended of a plurality of candidates is subjected to diversity sorting, wherein the diversity sorting is performed according to at least one of the following parameters of the information to be recommended of the plurality of candidates: the number of information from the same media account, the number of information from the same channel in the same media account, the number of non-premium information; and executing the recommendation operation corresponding to the user based on the result of the diversity sorting.

11. An electronic device, the electronic device comprising:

a memory for storing executable instructions;

A processor for implementing the artificial intelligence based information recommendation method of any one of claims 1 to 9 when executing executable instructions stored in the memory.

12. A computer readable storage medium storing executable instructions which when executed by a processor implement the artificial intelligence based information recommendation method of any one of claims 1 to 9.