CN111400603A

CN111400603A - Information pushing method, device and equipment and computer readable storage medium

Info

Publication number: CN111400603A
Application number: CN202010201835.9A
Authority: CN
Inventors: 刘毅; 饶君; 丘志杰; 张博
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Shenzhen Yayue Technology Co ltd
Priority date: 2020-03-20
Filing date: 2020-03-20
Publication date: 2020-07-10

Abstract

The invention provides an information pushing method, an information pushing device, information pushing equipment and a computer readable storage medium; the method comprises the following steps: when an information pushing instruction is obtained, according to the identity information of a user in the information pushing instruction, recalling information is extracted from a pushing information database; extracting characteristic information of the user according to the identity information, and acquiring historical category information of the user according to the identity information; the historical category information represents the category to which the information read by the user belongs; predicting category information to be pushed based on the feature information and the historical category information; the category information to be pushed represents the probability that the user is interested in each category; and screening the information to be pushed from the recall information according to the category information to be pushed, and sending the information to be pushed to a terminal of a user. By the method and the device, the personalized pushing degree of the user can be improved.

Description

Information pushing method, device and equipment and computer readable storage medium

Technical Field

The present invention relates to artificial intelligence technologies, and in particular, to an information pushing method, an information pushing apparatus, information pushing equipment, and a computer-readable storage medium.

Background

The information push refers to a process of screening out contents which may be interested by a user from an information database and then presenting the screened contents to the user. In the information pushing process, a large amount of information is generally recalled first, then the recalled information is subjected to rough screening, that is, information which may be interested by a user is primarily screened out, and then further the information obtained by the rough screening is subjected to fine screening to obtain an information set which is finally pushed to the user, so that the pushed information set fits the interest of the user.

In the related art, in order to increase the diversity of the information categories of the rough-row screening, a diversity queue can be added in the rough-row screening process, and the information in the queue has diversity and can be directly transmitted to the fine-row screening, so that the information pushed to the user has diversity; or the proportion of the recall information of different queues is distributed during rough screening, so that different types of recall information can be pushed to the user.

However, the competition degree between the diversity queue and other recall queues is low, and the pushed information is in fit with the interest of the user to a low degree, so that the personalized pushing degree for the user is low. Meanwhile, the proportion of the recalled information of different queues is difficult to change after the recall information is determined, so that the queue proportions of the information pushed by different users are the same, and the personalized pushing degree for the users is lower.

Disclosure of Invention

The embodiment of the invention provides an information pushing method, an information pushing device, information pushing equipment and a computer readable storage medium, which can improve the personalized pushing degree of a user.

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

the embodiment of the invention provides an information pushing method, which comprises the following steps:

when an information pushing instruction is obtained, according to the identity information of a user in the information pushing instruction, recalling information is extracted from a pushing information database;

extracting feature information of the user according to the identity information, and acquiring historical category information of the user according to the identity information; the historical category information represents the category to which the information read by the user belongs;

predicting category information to be pushed based on the feature information and the historical category information; the category information to be pushed represents the probability that the user is interested in each category;

and screening the information to be pushed from the recall information according to the category information to be pushed, and sending the information to be pushed to the terminal of the user.

An embodiment of the present invention provides an information pushing apparatus, including:

the recall module is used for extracting recall information from a pushed information database according to the identity information of a user in the information push instruction when the information push instruction is acquired;

the extraction module is used for extracting the characteristic information of the user according to the identity information and acquiring the historical category information of the user according to the identity information; the historical category information represents the category to which the information read by the user belongs;

the prediction module is used for predicting the category information to be pushed based on the characteristic information and the historical category information; the category information to be pushed represents the probability that the user is interested in each category;

and the information screening module is used for screening the information to be pushed from the recall information according to the category information to be pushed and sending the information to be pushed to the terminal of the user.

An embodiment of the present invention provides an information push apparatus, including:

the memory is used for storing an executable information pushing instruction;

and the processor is used for realizing the information push method provided by the embodiment of the invention when executing the executable information push instruction stored in the memory.

The embodiment of the invention provides a computer-readable storage medium, which stores an executable information pushing instruction and is used for causing a processor to execute the executable information pushing instruction so as to realize the information pushing method provided by the embodiment of the invention.

The embodiment of the invention has the following beneficial effects:

in the embodiment of the invention, when the information pushing equipment acquires the information pushing instruction, the recall information is extracted from the pushing information database according to the identity information of the user in the information pushing instruction, then the information pushing equipment extracts the feature information of the user according to the identity information and acquires the historical category information of the user according to the identity information, then the information pushing equipment predicts the category information to be pushed based on the feature information and the historical category information, and then the information to be pushed is screened out from the recall information according to the category information to be pushed, so that the information to be pushed is sent to the terminal of the user. Therefore, the information pushing equipment can predict which categories are more interested by the user according to the self attributes of the user and the historical reading records of the user, and then determine the information to be pushed for the user according to the categories which are interested by the user, so that the content which is not interested by the user can be pushed for the users according to the interests of the different users, and the personalized pushing degree of the user is improved.

Drawings

Fig. 1 is a schematic diagram of an alternative architecture of an information push system 100 according to an embodiment of the present invention;

fig. 2 is a schematic structural diagram of a server 200 in fig. 1 according to an embodiment of the present invention;

fig. 3 is a schematic flow chart of an alternative information pushing method according to an embodiment of the present invention;

FIG. 4 is a diagram illustrating obtaining historical category information according to an embodiment of the present invention;

fig. 5 is a schematic diagram of an alternative flow chart of an information pushing method according to an embodiment of the present invention;

fig. 6 is a schematic flow chart diagram three of an alternative information pushing method according to an embodiment of the present invention;

FIG. 7 is a schematic diagram of the present invention for constructing attribute feature vectors using feature information;

fig. 8 is a block diagram of a GRU unit provided in an embodiment of the present invention;

FIG. 9 is a diagram illustrating a method for determining class feature vectors according to an embodiment of the present invention;

fig. 10 is a schematic flow chart of an alternative information pushing method according to an embodiment of the present invention;

fig. 11 is a schematic diagram of an information pushing process provided by an embodiment of the present invention;

fig. 12 is a schematic diagram of information pushing in a glance according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail with reference to the accompanying drawings, the described embodiments should not be construed as limiting the present invention, and all other embodiments obtained by a person of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

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

In the description that follows, references to the terms "first", "second", and the like, are intended only to distinguish between similar objects and not to indicate a particular ordering for the objects, it being understood that "first", "second", and the like may be interchanged under certain circumstances or sequences of events to enable embodiments of the invention described herein to be practiced in other than the order illustrated or described herein.

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 invention belongs. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to be limiting of the invention.

Before further detailed description of the embodiments of the present invention, terms and expressions mentioned in the embodiments of the present invention are explained, and the terms and expressions mentioned in the embodiments of the present invention are applied to the following explanations.

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

The artificial intelligence technology is a comprehensive subject, and relates to the technologies of wide fields, the existing hardware level technology and the software level technology, and the artificial intelligence basic technology generally comprises a sensor, a special artificial intelligence chip, cloud computing, distributed storage, a big data processing technology, an operation/interaction system, electromechanical integration and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning, deep learning and the like.

2) Machine learning (Machine L earning, M L) is a multi-domain cross discipline, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. a special study on how a computer simulates or implements human learning behavior to acquire new knowledge or skills, reorganizes existing instruction structures to continuously improve its performance.

3) Deep Neural Networks (DNNs) are composed of a plurality of feedforward Neural networks and activation functions, and are widely applied to a plurality of scenarios such as information push, semantic understanding, semantic analysis, and the like.

4) Recurrent Neural Networks (RNNs) are a class of Neural networks with short-term memory. In the RNN, a neuron receives not only information of other neurons but also information of itself, and forms a network result having a loop.

5) A gated round-robin Unit (GRU) is a variant of the long-Short Term Memory network (L ong Short-Term Memory, L STM) that combines a forgetting Gate with an input Gate into a single update Gate.

6) The Stochastic Gradient Descent (SGD) is an optimal algorithm in Gradient Descent algorithms, and may also be referred to as a steepest Descent method. SGD is one of the simplest and oldest methods for solving unconstrained problems, and many effective gradient descent algorithms are improved or modified based on SGD. The SGD uses the negative gradient direction as the search direction, and the step size decreases and the progress slows as the search direction approaches the target value.

7) Collaborative Filtering (CF) refers to a method for pushing information according to the Collaborative effect of different user interests.

8) Collaborative filtering based on items (Item CF), the basic idea is to calculate the similarity between items in advance according to the historical preference data of all users, and then push other items similar to the items that the users like to the users.

9) Based on the collaborative filtering (User CF) of users, the basic idea is to perform similarity calculation on users according to the preference degrees of the users to the articles, so that the articles favored by the similar users are pushed to each other.

10) The sequence-based pushing refers to extracting a context relationship from a click sequence of a user, and then predicting information content which is likely to be interested by the user in the future according to the context relationship.

11) Based on content push, the basic idea is to mine the characteristics of the information content historically read by the user in the semantic dimension, and then calculate the similarity of the information content historically read by the user and other information content in the semantic dimension according to the characteristics, so as to recommend information for the user.

12) Feature extraction, which aims to extract features from raw data to the maximum extent that algorithms and models use these features.

13) Click Through Rate (CTR) is widely used in information recommendation, especially in advertisement recommendation. The CTR refers to the click arrival rate of the network advertisement (picture advertisement, text advertisement, keyword advertisement, ranking advertisement, video advertisement, etc.), that is, the actual number of clicks of the advertisement is divided by the display amount of the advertisement.

14) Bucketing, meaning dividing information into multiple sets, where each set can be thought of as a bucket. For example, when there are multiple categories of articles in the push information database, the articles in the food category may be regarded as a bucket, the articles in the current category may be regarded as a bucket, and so on. There are various ways of performing the bucket allocation, for example, there are ways of performing the bucket allocation according to the category of the information, ways of performing the bucket allocation according to the heat of the information, ways of performing the bucket allocation according to the type of recall, and ways of performing the bucket allocation according to other ways.

15) Static binning may be understood as setting a forced proportion to each bucket such that information in different buckets is presented according to the set proportion. For example, a 50% ratio is set for articles of the food category, a 20% ratio is set for articles of the current category, a 30% ratio is set for articles of the financial category, and the like. As another example, a 20% ratio is set for tentative recalls, a 30% ratio is set for social recalls, and a 50% ratio is set for precision recalls.

16) The cold start information refers to information content with low read times or low heat in the push information database, for example, a newly issued article is read less times and belongs to cold start information in general.

With the research and progress of artificial intelligence technology, the artificial intelligence technology has developed research and application in a plurality of fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned, autonomous, unmanned aerial vehicles, smart medical, smart customer service, and the like. It is believed that with the development of technology, artificial intelligence technology will find application in more fields and will play an increasingly important role.

Information push is an application of artificial intelligence technology. The information push refers to a process of screening out contents which may be interested by a user from an information database and then presenting the screened contents to the user. Information push can be divided into collaborative filtering based on articles, collaborative filtering based on users, push based on sequences, push based on contents, and the like according to a push manner.

At present, a commonly used information pushing method is to recall a large amount of information, for example, ten-thousand-level information, through a multi-recall layer, and then to roughly sort the recalled information by using a rough-sorting layer, for example, to sort out thousand-level content from the ten-thousand-level content, so that information that a user may be interested in can be preliminarily sorted out. And then, performing fine-line screening on the information obtained by the coarse-line screening by using a fine-line layer, for example, re-screening 10-order content from thousands-order content to obtain an information set which is finally pushed to the user, so that the pushed information set fits the interest of the user. Specifically, the multi-recall layer includes heuristic recalls (e.g., recalls according to user interests, content recalls), general precise recalls (e.g., category recalls, knowledge graph recalls), social recalls (e.g., friends are reading), precise recalls (e.g., recalls with DNN model, ItemC F, UserCF), and the like. Generally, to reduce the computation of the fine-ranking layer, the coarse-ranking layer filters the information content by using a CTR model with a far lower feature quantity and feature complexity than the fine-ranking layer, for example, reducing recalled ten-thousand content to thousand content by using a model with low feature complexity.

When the information content is screened by the coarse arrangement layer, the information content is scored by directly using the CTR model of the coarse arrangement layer, the information content with high score is extracted and transmitted to the fine arrangement layer, and in the process, the coarse arrangement layer does not have the variety of recall types. However, since the CTR model tends to select information content that is sufficiently exposed in the entire network, it is difficult to select cold start content that is underexposed, so that the timeliness of pushing information is insufficient. Meanwhile, the CTR model also tends to select content that is strongly interest-related, which may result in a high exposure duty of information content through accurate recall, e.g., the accurate recall duty exceeds 80%, etc. The precisely recalled information content is usually concentrated on some categories, so that the diversity of the pushed information is insufficient, i.e. the pushed information is single and homogeneous.

In order to increase the diversity of the rough row screening, a diversity queue can be introduced in the rough row screening process, or different recall queues are statically bucketed in the rough row screening process. The introduction of the diversity queue in the course of coarse screening means that a diversity queue whose information content satisfies diversity is added, and the information content in the queue is ensured to be directly transmitted to the fine-ranking layer, and the rest information content except the queue is still screened by adopting CTR model scoring. Static binning of different recall queues in the coarse sort screen refers to assigning different proportions for different recall types to ensure fairness of the various recall queues, e.g., 10% for tentative recalls, 20% for general precise recalls, 20% for social recalls and 50% for precise recalls. Thus, the information content screened out in the rough row has diversity.

However, the diversity queue introduced in the course of the rough-ranking screening does not compete with other recall queues, that is, the competition between the diversity queue and other recall queues is low, so that the click rate of the information content belonging to the queue is low, that is, the pushed information is not interesting for the user, and thus the personalized pushing effect for the user is poor. Meanwhile, the occupation ratio of the diversity queue directly transmitted to the fine ranking layer is also considered to be controlled, and no individuation exists, so that the occupation ratios of the diversity information contents pushed to users who like the diversity contents and users who do not like the diversity contents are the same, and the individuation is lacked in the information pushing process. For example, for an instant game, some users expect the pushed information to be concentrated on the strategy and tactical video of the game, while another user expects the pushed information to have not only the strategy and tactical video of the game, but also the animation surroundings of the game, news related to the game, and information content of other games similar to the game. When different recall queues are statically barreled in the rough-row screening, the proportion of recall information of different queues cannot be changed after the recall information is determined, so that the queue proportions of information pushed by different users are the same, and the personalized pushing effect on different users is lacked. Meanwhile, the information content in each recall queue can be continuously optimized in an iterative manner, the content and the category of the recall information can be changed before and after optimization, but the change cannot be updated in real time by the static bucket.

Therefore, the rough filtering in the related art lacks personalized recommendation for different users, namely the personalized push degree for the users is low.

The embodiment of the invention provides an information pushing method, an information pushing device, information pushing equipment and a computer readable storage medium, which can improve the personalized pushing degree of a user. An exemplary application of the information pushing apparatus provided in the embodiment of the present invention is described below, and the information pushing apparatus provided in the embodiment of the present invention may be implemented as various types of user terminals such as a smart phone, a tablet computer, and a notebook computer, and may also be implemented as a server. Next, an exemplary application when the information push apparatus implements the terminal and the server, respectively, will be described.

Referring to fig. 1, fig. 1 is an alternative architecture diagram of an information push system 100 according to an embodiment of the present invention, in order to support an information push application, a terminal 400 is connected to a server 200 through a network 300, where the network 300 may be a wide area network or a local area network, or a combination of the two. Wherein the server 200 is configured with a push information database 500.

Further, 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 providing basic cloud computing services such as a cloud service, a cloud database, cloud computing, a cloud function, cloud storage, a network service, cloud communication, a middleware service, a domain name service, a security service, a CDN, a big data and artificial intelligence platform, and the like. The terminal 400 may be, but is not limited to, a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, and the like. The terminal 400 and the server 200 may be directly or indirectly connected through wired or wireless communication, and the present invention is not limited thereto.

When the server 200 acquires the information push instruction, it extracts the recall information from the push information database 500 according to the identity information of the user in the information push instruction. Then, the server 200 obtains the feature information of the user according to the identity information, and extracts the history category information of the user according to the identity information. Next, the server 200 predicts category information to be pushed based on the feature information and the history category information, where the category information to be pushed represents a probability that the user is interested in each category, and then screens out the information to be pushed from the recall information according to the category information to be pushed, and sends the information to be pushed to the terminal 400 of the user. After receiving the information to be pushed sent by the server 200 and acquiring the information display instruction triggered by the user, the terminal 400 may present the information to be pushed on the display interface 410 of the terminal, thereby completing the information pushing process.

Referring to fig. 2, fig. 2 is a schematic structural diagram of a server 200 in fig. 1 according to an embodiment of the present invention, where the server 200 shown in fig. 2 includes: at least one processor 210, memory 250, at least one network interface 220, and a user interface 230. The various components in server 200 are coupled together by a bus system 240. It is understood that the bus system 240 is used to enable communications among the components. The bus system 240 includes a power bus, a control bus, and a status signal bus in addition to a data bus. For clarity of illustration, however, the various buses are labeled as bus system 240 in fig. 2.

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

The user interface 230 includes one or more output devices 231, including one or more speakers and/or one or more visual display screens, that enable the presentation of media content. The user interface 230 also includes one or more input devices 232, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

The memory 250 includes volatile memory or nonvolatile memory, and may include both volatile and nonvolatile memory. The non-volatile Memory may be a Read Only Memory (ROM), and the volatile Memory may be a Random Access Memory (RAM). The memory 250 described in embodiments of the invention is intended to comprise any suitable type of memory. Memory 250 optionally includes one or more storage devices physically located remotely from processor 210.

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

An operating system 251 including system programs for processing 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 processing hardware-based tasks;

a network communication module 252 for communicating to other computing devices via one or more (wired or wireless) network interfaces 220, exemplary network interfaces 220 including: bluetooth, wireless-compatibility authentication (Wi-Fi), and Universal Serial Bus (USB), etc.;

a display module 253 to enable presentation of information (e.g., a user interface for operating peripherals and displaying content and information) via one or more output devices 231 (e.g., a display screen, speakers, etc.) associated with the user interface 230;

an input processing module 254 for detecting one or more user inputs or interactions from one of the one or more input devices 232 and translating the detected inputs or interactions.

In some embodiments, the information pushing apparatus provided by the embodiments of the present invention may be implemented in software, and fig. 2 illustrates an information pushing apparatus 255 stored in the storage 250, which may be software in the form of programs and plug-ins, and includes the following software modules: recall module 2551, extraction module 2552, prediction module 2553, and information filtering module 2554, the functions of each of which will be described below.

In other embodiments, the information pushing apparatus provided in the embodiments of the present invention may be implemented in hardware, and as an example, the information pushing apparatus provided in the embodiments of the present invention may be a processor in the form of a hardware decoding processor, which is programmed to execute the information pushing method provided in the embodiments of the present invention, for example, the processor in the form of the hardware decoding processor may employ one or more Application Specific Integrated Circuits (ASICs), DSPs, Programmable logic devices (P L D, Programmable L analog devices), Complex Programmable logic devices (CP L D, Complex Programmable logic devices L analog devices), Field Programmable Gate Arrays (FPGAs), or other electronic elements.

Illustratively, an embodiment of the present invention provides an information push apparatus, including:

the memory is used for storing an executable information pushing instruction;

In the following, the information push method provided by the embodiment of the present invention will be described in conjunction with exemplary applications and implementations of the information push apparatus provided by the embodiment of the present invention.

Referring to fig. 3, fig. 3 is a first alternative flow chart of an information pushing method according to an embodiment of the present invention, which will be described with reference to the steps shown in fig. 3.

S101, when the information push instruction is obtained, recalling information is extracted from a push information database according to the identity information of the user in the information push instruction.

The embodiment of the invention is realized under the scene of pushing information for the user according to the interest and reading history of the user. The information pushing process is triggered by an information pushing instruction, when the information pushing equipment acquires the information pushing instruction, the information pushing equipment indicates that information to be pushed needs to be determined for a user at the moment, the information pushing equipment can analyze the identity information of the user from the information pushing instruction, information extraction is carried out on the user from a pushing information database according to the identity information of the user, and all the obtained information is recall information.

The information push instruction may be generated by the information push apparatus itself, or may be sent by the terminal of the user. Further, the information pushing instruction may be generated by the information pushing device after the preset pushing time is reached, for example, more users usually click and read the pushing information between 12:00 and 13:00, so that the information pushing device may generate the information pushing instruction at 11:50 of every day to start the information pushing process, thereby obtaining the information to be pushed and completing the information pushing process, so that the user can read the information content of interest at 12: 00. In addition, when the user has a need of reading information content of interest, an information pushing instruction can be sent to the information pushing device through the terminal, for example, the user can click a refresh button on a display interface displaying the pushed information, so that the information pushing device is informed to start information pushing.

It can be understood that the identity information of the user in the information push instruction may be an account number, a nickname, and the like of the user, and may also be other identification information that can indicate the identity of the user like the information push device.

In the embodiment of the invention, the information pushing equipment can extract the recall information from the pushed information database through the multi-channel recall queue. In the multi-way recall queue, a tentative recall queue, a general accurate recall queue, a social recall queue, an accurate recall queue and the like can be provided, and the recall queues can index interest tags of users, reading histories of friends, historical reading categories of users, and even popular content in a network and the like by utilizing identity information of the users, so as to obtain recall information.

It is understood that each piece of recall information has a corresponding information category, i.e., the recall information is composed of information belonging to different categories.

It should be noted that the pushed information database in the embodiment of the present invention is a database composed of information that can be pushed to a user, and in the pushed information database, there may be information contents that have already been read by the user, and also information contents that have not been read by the user.

S102, acquiring characteristic information of a user according to the identity information, and acquiring historical category information of the user according to the identity information; the historical category information represents the category to which the information read by the user belongs.

The information pushing device extracts various sub-feature information of the user from a database storing the feature information according to the identity information, and then uses all the extracted sub-feature information to form the feature information, wherein each sub-feature information represents an attribute feature or an interest feature of the user. Because it is difficult to cover the historical behavior characteristics of the user more comprehensively according to the self-owned attribute characteristics of the user, the information pushing device needs to obtain the category to which the information read by the user belongs according to the identity information of the user, that is, obtain the historical category information of the user, in addition to the characteristic information of the user, so as to predict the category which the user may be interested in subsequently by combining the characteristic information and the historical category information of the user.

It can be understood that, because a single piece of feature information is hard to show the attribute characteristics of the user, and the feature information composed of a large amount of sub-feature information can basically describe the attribute characteristics of the user, in the embodiment of the present invention, the information push device uses all the obtained sub-feature information to compose the feature information of the user. In other words, the feature information may be understood as a representation of the user.

In the embodiment of the present invention, the attribute feature refers to the self-owned basic feature of the user, for example, gender: male, area: shanghai, etc., and the interest features are generated based on the tags of the user, or based on the interests selected by the user in the selection page, e.g., interests: sketch, food, historical reading category: health preserving, cancer prevention and the like.

It should be noted that the history category information acquired by the information pushing device refers to a preset number of sub-category information acquired from the current time point in the reverse order of time. The preset number may be set according to actual situations, for example, to be 30 or to be 50, and the embodiment of the present invention is not limited herein.

Further, it is highly probable that the preset number of category information do not belong to the same historical time point, but belong to one or more historical time points, and each time point has corresponding sub-category information.

For example, as shown in fig. 4, the current time point 4-1 is 12:00, and the user reads the information content of the food category 4-2, the information content of the photographing category 4-3, and the information content of the favorite category 4-4 in sequence before 12:00, respectively. When the preset number is 2, the information pushing device sequentially acquires the category information from 12:00 according to the reverse direction of time, namely acquiring the budding pet category 4-4 first and then acquiring the photographic category 4-3, and further acquiring the historical category information.

S103, predicting to-be-pushed category information based on the feature information and the historical category information; the category information to be pushed represents the predicted probability that the user is interested in each category.

After obtaining the feature information and the historical category information of the user, the information pushing device may conjecture the probability that the user is interested in the information in each category next by combining various attribute features and interest features of the user and the category of the information read by the user history, that is, predict the category information to be pushed, so as to realize the subsequent pushing of the information for the user according to the interest of the user.

In the embodiment of the invention, the attribute features in the feature information correspond to a part of categories, the interest features also correspond to a part of categories, the categories corresponding to the feature information are likely to have coincidence with the categories in the history category information, and the coincidence can indicate the interest of the user to a certain extent. For example, when the feature information of the user corresponds to a food class and the number of times of occurrence of the food class in the history class information is large, the information pushing device may consider that the user is interested in the information of the food class, that is, the probability of interest for the food class is large, for example, 0.6; when the feature information of the user corresponds to a current affair, but the number of times of occurrence of the current affair in the history category information is small, the information pushing device may consider that the probability of interest of the user in the information of the current affair is medium, for example, 0.35; when there is no health preserving class in the category corresponding to the feature information of the user and there is almost no health preserving class in the history category information, the information pushing device may consider that the user is not interested in the information of the health preserving class, that is, the probability of interest in the information of the health preserving class is small, for example, 0.15. Thus, the information pushing device can predict the probability of interest of the user for each category.

Further, in some embodiments of the present invention, in order to calculate the probability that the user is interested in each category more accurately, the information pushing device may construct an attribute feature vector according to the feature information of the user, and construct a category feature vector according to the historical category information of the user, so as to predict the attribute feature vector and the category feature vector by using a trained category prediction model, or predict a vector obtained by splicing the attribute feature vector and the category feature vector by using the trained category prediction model, so as to obtain the probability that the user is interested in each category.

It should be noted that, because the predicted probability of interest of the user for each category can be described to some extent, when the probability of interest exceeds a probability threshold, for example, 0.6, it is described that the user is interested in the push information of the category, and the user is likely to open the push information belonging to the category corresponding to the probability of interest by clicking, otherwise, it is described that the user is not interested in the push information of the category, and the user hardly opens the push information belonging to the category.

Illustratively, when the probability of the interest of the user in the food class reaches 80%, it indicates that the user is more interested in the push information of the food class, so that the user is more likely to click to open the push information of the food class. When the probability of the interest of the user in the entertainment eight diagrams is 20%, the user is not interested in the push information of the entertainment eight diagrams, and the push information belonging to the entertainment eight diagrams is hardly opened.

S104, according to the category information to be pushed, information to be pushed is screened from the recall information, and the information to be pushed is sent to a terminal of a user.

After the information pushing device obtains the category information to be pushed, the information to be pushed can be screened out from the information which belongs to each information category in the recalling information according to the category information to be pushed, namely the probability of interest of the user for each information category. Further, the information pushing device may obtain more information for the information category with a higher probability of user interest, and obtain less information for the information category with a lower probability of user interest, so that the information to be pushed is matched with the interest of the user. The information to be pushed by different users is different in category information and further different in information to be pushed, so that the information pushing equipment realizes personalized pushing for different users.

It can be understood that the information pushing device picks out categories of the category information to be pushed, of which the probability exceeds the probability threshold, and then extracts only information belonging to the categories from the recall information, and the information can be acquired from the recall information by proportional to the probability of each category in the category information to be pushed.

In the embodiment of the invention, when the information pushing equipment acquires the information pushing instruction, the recall information is extracted from the pushing information database according to the identity information of the user in the information pushing instruction, then the information pushing equipment extracts the feature information of the user according to the identity information and acquires the historical category information of the user according to the identity information, then the information pushing equipment predicts the category information to be pushed based on the feature information and the historical category information, and then the information to be pushed is screened out from the recall information according to the category information to be pushed, so that the information to be pushed is sent to the terminal of the user. Therefore, the information pushing equipment can predict which categories are more interested by the user according to the self attributes of the user and the historical reading records of the user, and then determine the information to be pushed for the user according to the categories which are interested by the user, so that the content which is respectively interested by different users can be pushed for the different users according to the interests of the different users, and the personalized pushing degree of the user is improved.

It should be noted that, in some embodiments of the present invention, when the information pushing device obtains the information pushing instruction, a process of extracting the feature information of the user and the historical category information of the user according to the identity information of the user in the information pushing instruction, predicting the category information to be pushed based on the feature information and the historical category information, and then a process of extracting the recall information according to the identity information may be performed, that is, S102 to S103 are performed first, then S101 is performed, and finally S104 is performed. In other embodiments of the present invention, the processes of S102-S103 and the process of S101 may also be performed simultaneously.

In some embodiments of the present invention, when the information push instruction is obtained, the retrieving information is extracted from the push information database according to the identity information of the user in the information push instruction, that is, a specific implementation process of S101 may include: S1011-S1013, as follows:

and S1011, when the information push instruction is acquired, extracting a plurality of initial sub-information from the push information database according to the identity information.

When the information pushing equipment extracts the recall information, initial sub-information which is not processed is extracted from a pushed information database through a multi-channel recall queue according to the identity information. Since each recall queue of the multi-way recall queue can obtain at least one initial sub-message, the information push device can extract a plurality of initial sub-messages from the push information database.

It is understood that, in some embodiments of the present invention, the number of the plurality of initial sub information may reach ten thousand, for example, 10000 initial sub information are extracted.

And S1012, classifying the plurality of initial sub information according to the class label of each initial sub information to obtain an initial sub information set corresponding to each class.

Each piece of information in the pushed information database is marked with a category label when being generated, so that all initial sub-information extracted by the information pushing equipment has the category label of the information pushing equipment, at the moment, the information pushing equipment can read the category label of each piece of initial sub-information and divide the initial sub-information with the same category label into the same category, namely, the category label is used for dividing the extracted sub-information into categories. After completing the classification of all the initial sub-information, the information pushing device may form an initial sub-information set with the initial sub-information belonging to the same class, so that the information pushing device may obtain an initial sub-information set corresponding to each class.

And S1013, forming recall information by utilizing all the initial sub-information in the initial sub-information set.

The information push equipment uses all initial sub-information in each initial sub-information set to form recall information, in other words, each initial sub-information in the recall information is subjected to bucket division and has a category to which the initial sub-information belongs, so that the information push equipment can directly acquire information in each bucket when screening information to be pushed from the recall information.

It can be understood that the information pushing apparatus may directly use all the initial sub information in each initial sub information set to form the recall information, or may perform processing such as sorting on the initial sub information in each initial sub information set, and use the processed initial sub information to form the recall information.

In the embodiment of the invention, the information pushing equipment can extract a plurality of initial sub-information from the pushing information database according to the identity information, then perform category division on the plurality of initial sub-information according to the category label of each initial sub-information to obtain the initial sub-information set corresponding to each category, and finally, form the recall information by using all the initial sub-information in the initial sub-information set. Therefore, the information pushing equipment can obtain the recalled information which passes through the bucket.

Referring to fig. 5, fig. 5 is a schematic view illustrating an optional flow chart of an information pushing method according to an embodiment of the present invention. In some embodiments of the present invention, after performing category division on a plurality of pieces of initial sub information according to a category label corresponding to each piece of initial sub information to obtain the initial sub information corresponding to each category, before composing the recall information by using all pieces of initial sub information in the set of initial sub information, that is, after S1012 and before S1013, the method may further include: s1014, the following steps are carried out:

s1014, sequencing all initial sub information in the initial sub information set to obtain a sub information set corresponding to each category.

After the information pushing device obtains the initial sub-information sets corresponding to each category, the information pushing device may sort each initial sub-information in the initial sub-information sets, and form a sub-information set by using the sorted initial sub-information. In this way, the information pushing device can obtain the sub-information sets corresponding to each category one by one. Accordingly, after the information pushing apparatus obtains the sub-information set, the process of originally composing the recall information by using all the initial sub-information in the initial sub-information set, that is, the process of S1013, is changed to compose the recall information by using all the sorted initial sub-information in the sub-information set.

It is understood that the information pushing device may use the CTR model to sort the score of the initial sub-information, and may also sort the initial sub-information according to the popularity, the release time, and the like of the initial sub-information. The specific sorting manner may be set according to actual situations, and the embodiment of the present invention is not limited herein.

In the embodiment of the invention, the information pushing equipment can sort the initial sub-information in the initial sub-information set to obtain the sub-information set, and then the initial sub-information sorted in the sub-information set is utilized to form the recall information. In this way, when the information pushing device filters the information to be pushed, the information to be pushed can be filtered from the sorted initial sub-information of the recall information, so that the initial sub-information with the front order can be preferentially filtered.

In some embodiments of the present invention, the sorting the initial sub information in the initial sub information set to obtain the sub information set corresponding to each category, that is, the specific implementation process of S1014 may include: s1014a-S1014b, as follows:

s1014a, acquiring the content characteristics of the initial sub information in the initial sub information set, and calculating the ranking score of the initial sub information by using the content characteristics.

In the embodiment of the invention, the information pushing equipment scores the initial sub-information so as to finish sequencing. At this time, the information pushing device may first obtain the content characteristics for each piece of initial sub information, then obtain the trained CTR model from its own storage space, and then input the content characteristics of each piece of initial sub information into the CTR model for calculation, where what the CTR model outputs is the ranking score of each piece of initial sub information.

It should be noted that the content feature of the initial sub information includes a title of the initial sub information, a tag of a user who has viewed the initial sub information, and the like, and of course, the content feature may also have other features capable of describing the characteristics of the initial sub information, and the embodiment of the present invention is not limited herein.

S1014b, sorting the initial sub-information by using the sorting fraction to obtain sorted initial sub-information, and forming a sub-information set by using the sorted initial sub-information.

After the information pushing device obtains the ranking score, the information pushing device may rank each initial sub-information by using a principle that the ranking score is from high to low, so as to obtain the ranked initial sub-information. Then, the information pushing device composes a sub information set by using the sorted initial sub information.

In the embodiment of the invention, the information pushing equipment can calculate the sorting score according to the content characteristics of the acquired initial sub-information, and then sort the initial sub-information according to the sorting score, so that the sorted initial sub-information is obtained, and further, the sub-information set is obtained. Therefore, when the information pushing equipment filters information to be pushed from the recall information in the follow-up process, the initial sub-information with the highest ranking score in each sub-information set can be selected preferentially, so that the personalization degree of information pushing is further improved.

In some embodiments of the present invention, predicting category information to be pushed based on the feature information and the historical category information, that is, a specific implementation process of S103 may include: S1031-S1033, as follows:

and S1031, constructing attribute feature vectors of the users by using the feature information.

Because the feature information is a set formed by a plurality of pieces of sub-feature information, when the information pushing device uses the feature information to construct the attribute feature vector of the user, each piece of sub-feature information is firstly converted into a vector form, namely, converted into a sub-feature vector, and then the sub-feature vector is used to construct the attribute feature vector of the user, so that the category information to be pushed is constructed by using the attribute feature vector subsequently.

In some embodiments of the present invention, the information pushing device may directly splice the sub-feature vectors corresponding to each piece of sub-feature information into an attribute feature vector, so that the attribute feature vector can describe the attribute features and interest features of the user comprehensively.

Furthermore, some features possibly existing in the feature information have little influence on the process of determining the category information to be pushed, for example, a region where the user is located, and the like, so that the information pushing device can distinguish the importance degree of each piece of sub-feature information, and construct an attribute feature vector according to the more important sub-feature information. Still further, the combined result of some sub-feature information in the feature information can also affect the category information to be pushed, for example, a user with the combined characteristics of < mom, sensitive physique baby > may be more interested in an article of the type that introduces baby and infant supplementary food safety. Therefore, the information pushing device can mine the combination information among the features, and obtain the attribute feature vector according to the combination information, so that the attribute feature vector can better describe the self attribute of the user, and further the degree of personalized recommendation for the user is improved. For example, in some embodiments of the present invention, the information pushing device may perform self-attention calculation on a vector obtained by stitching with the sub-feature vectors to identify the importance degree of each sub-feature information, and mine more important combination information among the sub-feature information to obtain the attribute feature vector.

S1032, determining a category feature vector of the user according to the history category information; the category feature vector can represent the characteristics of the historical category information.

Because it is difficult to cover the historical behavior characteristics of the user more comprehensively according to the own attribute characteristics of the user, the information pushing device needs to construct a category characteristic vector according to the category to which the information read by the user belongs, that is, the historical category information, in addition to obtaining the attribute characteristic vector of the user, so as to describe the historical behavior characteristics of the user by using the category characteristic vector. In other words, the category feature vector is an abstract generalization of the characteristics that the user's historical category information has.

For example, the information pushing device can determine the category feature vector by using L STM, RNN and other models so as to embody the time sequence relation in the historical category information.

It should be noted that the category feature vector is a characteristic of the represented historical category information, and may be a category in which the reading times or reading frequency of the user in a recent period is high, a reading period of the user for different categories, or a category in which the user often reads near a special time node, for example, the user often reads gourmet articles on weekends, and the user often reads yearly-popular introduction articles before and after spring festival, and the like. Of course, the characteristics of the historical category that can be characterized by the category feature vector are not limited to these, and may represent abstract generalizations of other characteristics.

And S1033, predicting the category information to be pushed by using the attribute feature vector and the category feature vector.

After obtaining the attribute feature vector and the category feature vector of the user, the information pushing device may splice the attribute feature vector and the category feature vector into a vector capable of describing the attribute feature and the historical behavior feature of the user at the same time, predict the interest probability of the user for each category in the recall information from the vector, and use the interest probability of the user for each category as the category information to be pushed.

It should be noted that the execution order of constructing the attribute feature vector and determining the category feature vector may be interchanged, that is, in other embodiments of the present invention, S1032 may be executed first, then S1031 is executed, and finally S1033 is executed, or S1031 and S1032 may be executed at the same time, and finally S1033 is executed.

In the embodiment of the invention, the information pushing equipment can construct the attribute feature vector by using the characteristic information, determine the category feature vector by using the historical category information, and then predict the category information to be pushed by combining the attribute feature vector and the historical category vector. Therefore, the information pushing equipment can accurately calculate the interest probability of the user to each category, and improves the accuracy of the personalized pushing of the user on the basis of realizing the personalized pushing of the user.

Referring to fig. 6, fig. 6 is a third optional flowchart of the information pushing method according to the embodiment of the present invention. In some embodiments of the present invention, the feature information comprises one or more sub-feature information; the constructing of the attribute feature vector of the user by using the feature information, that is, the specific implementation process of S1031, may include: s1031a-S1031d, as follows:

and S1031a, performing vector transformation on each piece of sub-feature information in the one or more pieces of sub-feature information to obtain one or more sub-feature vectors.

When the information pushing equipment constructs the attribute feature vector, each piece of sub-feature information in the feature information is converted into a vector, and therefore a sub-feature vector corresponding to each piece of sub-feature information is obtained. Since the feature information has one or more sub-feature information, the information pushing device may obtain one or more sub-feature vectors.

It can be understood that the information pushing device may determine the sub-feature vector by encoding each piece of sub-feature information and using the encoded result as the sub-feature vector, or by presetting a corresponding relationship between the sub-feature information and the sub-feature vector.

Illustratively, when some sub-feature information of the user is male, the information pushing device may encode it as (0.4, 0.1, 0.7); when another piece of sub-feature information of the user is shanghai, it may be encoded as (0.1, 0.1, 0.3).

And S1031b, splicing the intermediate feature vector by using one or more sub-feature vectors.

S1031c, performing multiple self-attention calculations on the intermediate feature vector to obtain multiple attention parameters corresponding to the intermediate feature vector; wherein the attention parameter characterizes the degree of importance of the sub-feature information.

And the information pushing equipment splices all the obtained sub-feature vectors according to the sequence of the sub-feature information, and records the splicing result as an intermediate feature vector. Then, the information pushing device performs multiple self-attention calculations on the intermediate feature vector by using a multi-head self-attention mechanism, wherein each self-attention calculation corresponds to one head in the multi-head self-attention mechanism to obtain one attention parameter, and therefore, the intermediate feature vector corresponds to multiple attention parameters.

It should be noted that the purpose of the multi-head self-attention machine is to analyze the intermediate feature vector, so as to find out a part with a larger importance degree in the intermediate feature vector, where the part may be one sub-feature vector or a combination of several sub-feature vectors, and therefore, the attention parameter obtained by the multi-head self-attention machine is capable of characterizing the importance degree of the sub-feature information.

When the self-attention calculation is carried out, firstly, the intermediate feature vector is transposed to obtain a transposed intermediate feature vector, then, the product of the intermediate feature vector and the transposed intermediate feature vector is used, compared with the evolution result of the dimension of the intermediate feature vector, the ratio result is input into softmax, finally, the output result of the softmax is multiplied by the intermediate feature vector, and the obtained product is the attention parameter of one head.

For example, the embodiment of the present invention provides a calculation formula of an attention parameter of a head, that is, a formula of self-attention calculation, see formula (1):

wherein, Q is a Query (Query), K is a Key (Key), V is a Value (Value), and Attention is an Attention parameter. In the embodiments of the present invention, X ═ X₁,x₂,…,x_n]As intermediate feature vectors, d_kIs the dimension of X. When the information pushing device knows the intermediate eigenvector and its dimension, the above parameters can be substituted into equation (1) to calculate the Attention parameter Attention of one head.

Further, in the embodiment of the present invention, specifically, when calculating the attention parameters of a plurality of heads, the calculation is performed according to the formula (1) for each head, for example, the formula for the attention parameter of the ith head can be referred to as the formula (2):

head_i＝Attention(QW_i ^Q,KW_i ^K,VW_i ^V) (2)

wherein Q ═ W ═ V ═ X, X ═ X₁,x₂,…,x_n]Is an intermediate feature vector, W_i ^Q、W_i ^K、W_i ^QIs a trained parameter for the ith head in a multi-head self-attention mechanism.

And S1031d, splicing the plurality of attention parameters to obtain attribute feature vectors.

The information pushing equipment sequentially splices the calculated attention parameters to obtain a splicing result, then multiplies the splicing result by the trained overall parameters in the multi-self-attention mechanism, and the obtained product result is the attribute feature vector.

Illustratively, when the attention parameter of the ith head in the multi-attention mechanism is head_iThen, the attribute-property vector can be expressed as U ═ head₁,head₂,head₃,…,head_n]W^oWherein n is the total number of heads in the multi-head self-attention mechanism, W^oIs a well-trained overall parameter.

The characteristic information 7-1 includes sub-characteristic information gender 7-11, age 7-12, region 7-13, tag 7-14 and the like, input vector embedding layer 7-2 is used for vector conversion, then the converted sub-characteristic vectors enter self-attention layer 7-3 after splicing, self-attention calculation is carried out for multiple times, multiple attention parameters are obtained, and finally the information pushing device splices the final attribute characteristic vector by the multiple attention parameters.

In the embodiment of the invention, the information pushing equipment can convert each piece of sub-feature information into the sub-feature vectors, then the sub-feature vectors are used for splicing the intermediate feature vectors, then the attention calculation is carried out on the intermediate feature vectors to obtain the attention parameters, and finally the attention parameters are spliced to obtain the attribute feature vectors. Therefore, the information pushing equipment can construct the attribute feature vector so as to construct the category information to be pushed by using the attribute feature vector subsequently.

In some embodiments of the invention, the historical category information includes subcategory information corresponding to each historical point in time; determining the category feature vector of the user according to the history category information, that is, the specific implementation process of S1032 may include: s1032a-S1032c, as follows:

s1032a, carrying out vector conversion on the sub-category information to obtain a sub-category vector.

S1032b, predicting a history category vector corresponding to each history time point, using the sub-category vector corresponding to each history time point and the history category vector of the previous time point of each history time point.

The information pushing device has two inputs when predicting the history category vector of each time point, wherein one input is the predicted history category vector of the last time point of each time point. That is, the historical category vector is not only related to the sub-category vector, but also to the historical state of itself. It is understood that for an initial historical time point without a previous historical time point, an initial state may be set for the initial historical time point, that is, the initial state is set as a historical category vector of the previous time point of the initial time point.

In some embodiments of the invention, the information pushing device may utilize the GRU unit to predict the history category vector for each history time point. Fig. 8 is a block diagram of a GRU unit provided by an embodiment of the present invention, and as can be seen from fig. 8, the GRU unit has two inputs, one being x_iI.e. the sub-category vector, one is h (t-1), i.e. the historical category vector at the last time point, and finally the output is h_tI.e. a history category vector for each history time point. Where h is the main operation function in the GRU, z is update gating, and r is reset gating. When the history category vector at the time t is predicted by the information pushing equipment, the updating gating z at the time 1 and the time t is firstly used_tMaking difference, multiplying the difference result with the history category vector corresponding to the last time point of the t-1 moment, namely the t moment, to obtain a first parameter part, and then using the updated gating z of the t moment_tAnd a category vector calculation parameter h 'calculated in advance at time t'_tAnd finally, adding the first parameter part and the second parameter part to obtain the historical category vector at the time t.

For example, the embodiment of the present invention provides a calculation formula of a history category vector at time t, see formula (3):

h_i＝GRU(x_i,h_i-1)＝(1-z_t)·h_i-1+z_t·h′_i(3)

wherein, the category vector calculation parameter h at the time t_i' can be calculated from equation (4) as follows:

wherein W, U, b are all parameters that have been trained.

While the update gating for time t is z_tAnd reset gating r at time t_tAlso, formula (5):

[z_t,r_i]＝σ([W_z,W_r]X_t+[U_z,U_r]h_t-1+[b_z,b_r]) (5)

in these equations, W, U, b is a trained parameter.

S1032c, the history category vector corresponding to the history time point closest to the current time point is set as the category feature vector.

After obtaining the history category vector of each history time point, the information pushing device then takes the history category vector corresponding to the last history time point of all the history time points, that is, the history time point closest to the current time point, as the final category feature vector. Thus, the information pushing device abstracts and summarizes the category feature vector related to time from the historical category information.

For example, the embodiment of the present invention provides a schematic diagram for determining the category feature vector, see fig. 9, categories 9-11, categories 9-12, categories 9-13, and categories 9-14, until category 9-1n is the subcategory information, and the subcategory information is arranged according to the time sequence of the historical time point. Taking the initial state 9-21 as the history category vector of the last time point of the GRU unit at the initial history time point, after the initial state 9-21 and the category 9-11 enter the GRU unit at the initial history time point, the history category vector 9-22 at the initial history time point can be obtained, and further, the history category vector corresponding to the initial history time point (namely, 9-22, 9-23, 9-24, 9-25 … …) can be obtained at each history time point until the history category vector 9-2n at the history time point closest to the current time point is obtained, at this time, 9-2n is the category feature vector determined by the information pushing device.

In the embodiment of the invention, the information pushing equipment can convert the subcategory information into subcategory vectors, then the subcategory vector of each historical time point and the historical category vector output at the last time point are used for predicting the historical category vector of each historical time point, and finally the historical category vector of the historical time point closest to the current time point is used as the category feature vector. In this way, the information pushing device can obtain the category feature vector representing the characteristics of the historical category information, so that the category information to be pushed is constructed by using the category feature vector in the following process.

In some embodiments of the present invention, predicting category information to be pushed based on the attribute feature vector and the category feature vector, that is, a specific implementation process of S1033 may include: s1033a-S1033b, as follows:

and S1033a, splicing the attribute feature vectors and the category feature vectors to obtain the user feature vectors.

After the information pushing equipment obtains the attribute feature vector for describing the attribute features of the user and the category feature vector for describing the category features of the pushing information read by the user in the past, the two vectors can be spliced together to obtain the user feature vector capable of uniformly summarizing the attribute features and the historical behavior features of the user.

And S1033b, performing category information prediction on the user feature vector by using the trained category prediction model to obtain category information to be pushed.

The information pushing device obtains the trained category prediction model from the storage space of the information pushing device, then inputs the user characteristic vector into the category prediction model so as to predict the category information of the user characteristic vector through the category prediction model, and then the result output by the category prediction model is the category information to be pushed.

It is understood that, in the embodiment of the present invention, the category prediction model may be a softmax model, and may also be another machine learning model, for example, a deep neural network model, and the embodiment of the present invention is not limited herein.

Further, before the information pushing device predicts the category information, the category information model is trained, where the training of the category information model may be performed in a supervised training manner or in an unsupervised training manner, and the embodiment of the present invention is not limited herein.

When the category prediction model is trained by using a supervised training mode, firstly, an attribute feature vector and a category feature vector of a user are constructed according to the method, then, the user feature vector is constructed, then, the user feature vector is input into the category prediction model to obtain a prediction output, the category information which is actually read by the user is used as a supervision item, then, the supervision item and the prediction output are used for calculating a loss value, and therefore, the loss value is used for reversely adjusting parameters in the category prediction model until the trained category prediction model is obtained.

In the embodiment of the invention, the information pushing equipment can splice the user characteristic vector by using the attribute characteristic vector and the category characteristic vector, and then predict the category information of the user characteristic vector by using a trained category prediction model to obtain the category information to be pushed. Therefore, the information pushing equipment can accurately predict the interest degree of the user to each category according to the attribute feature vector and the category feature vector of the user, so as to further improve the accuracy degree of personalized recommendation for the user.

Referring to fig. 10, fig. 10 is a schematic view illustrating an optional fourth flow chart of the information pushing method according to the embodiment of the present invention. In some embodiments of the present invention, the screening of the information to be pushed from the recall information according to the category information to be pushed, that is, the specific implementation process of S104 may include: S1041-S1043, as follows:

s1041, calculating the information extraction quantity aiming at each sub information set in the recall information according to the category information to be pushed.

Since the category information to be pushed is the probability of interest of the user for each category, the information pushing device can directly use the probability of interest of the user for a certain category as the number ratio of the category, and then multiply the total number of the information to be acquired and the number ratio of the category to calculate the number of the information to be extracted for the category. After the above operation is completed for each category, the information extraction number of the sub information set corresponding to each category in the recall information can be obtained.

Exemplarily, when the category information to be pushed is C ═ C₁,c₂,…,c_nDenotes time (c)₁+c₂+…+c_n1), if the total number of sub information in the recall information is T, the basis is T_i＝T×c_iThe number of information extractions T ═ T can be calculated for each category₁,t₂,…,t_n}。

S1042, extracting the sub information of each sub information set according to the information extraction quantity to obtain middle sub information.

The information pushing equipment extracts the sub information according to the information extraction quantity of each category and the information extraction from the sub information set corresponding to each category in the recall information, and then takes all the extracted sub information of each category as the middle sub information.

It can be understood that, when the information pushing device extracts the sub information from each sub information set, the sub information may be extracted randomly, or extracted according to the ranking score of each sub information in the sub information set, or extracted according to an extraction manner, and the embodiment of the present invention is not limited herein.

And S1043, performing fine-arranging processing on the intermediate sub-information to obtain information to be pushed.

After the information pushing device obtains the intermediate sub-information, the intermediate sub-information needs to be continuously sent to the fine ranking layer, so that the intermediate sub-information is further screened and sorted through the fine ranking layer, and a result output by the fine ranking layer is used as final information to be pushed.

It is to be understood that the refinement processing may be implemented by using a CTR model of a refinement layer, or by using a model such as a deep neural network, and the embodiment of the present invention is not limited herein.

In the embodiment of the invention, the information pushing equipment can extract the sub-information from each sub-information set in the recall information according to the category information to be pushed to obtain the middle sub-information, and input the middle sub-information into the fine ranking layer for further processing. Therefore, the information pushing equipment completes the process of acquiring the information according to the interest degree of the user in each category, so that each piece of sub-information is more interesting to the user, and the personalized pushing degree for the user is improved.

In this embodiment of the present invention, extracting sub information from each sub information set according to the information extraction quantity to obtain intermediate sub information, that is, a specific implementation process of S1042 may include: s1042a-S1042b, as follows:

s1042a, starting the acquisition of the sub information from the sub information with the highest sorting score in each sub information set until the acquired information is extracted into sub information.

S1042b, using the acquired sub information as intermediate sub information.

The information pushing equipment extracts the front information in each sub information set according to the principle that the sorting scores are from high to low to obtain the sub information, and then all the obtained sub information is jointly used as the middle sub information so as to be convenient for fine sorting processing of the middle sub information.

In the embodiment of the invention, the information pushing equipment can obtain the front information and extract a number of pieces of sub information from each sub information set, so as to obtain the middle sub information. Therefore, the intermediate sub-information acquired by the information pushing device is higher in ranking score in each sub-information set, and the user is more likely to be interested, so that the personalized recommendation degree for the user is further improved.

In some embodiments of the present invention, before the information to be pushed is screened from the recall information according to the information to be pushed and is sent to the terminal of the user, that is, before S104, the method may further include: s105, the following steps are carried out:

s105, acquiring cold start information from a push information database; the cold start information is information that the number of times of reading is lower than a preset number threshold.

The information pushing device can obtain cold start information from the pushing information database besides the recall information from the pushing information database. The cold start information is information of a small number of exposures, that is, information of a number of times of reading by a person that is lower than a preset number threshold. This part of the information is most likely newly generated information, e.g. newly written articles, newly published songs, etc., or information that has already been generated and rarely read by people. In other words, the cold start message has a certain newness, and the characters with newness occupy not too much proportion in the recall message due to low reading times and hotness. Therefore, in order to enhance the pushing of the cold start information, the information pushing device may set a specific recall queue for the cold start information to recall the cold start information from the pushed information database. Then, the information pushing device screens out the information to be pushed from the recall information according to the category information to be pushed, and sends the information to be pushed to the terminal of the user, that is, the specific process of S104 will be correspondingly changed into: and the information pushing equipment screens out the information to be pushed from the recall information and the cold start information according to the category information to be pushed, and sends the information to be pushed to the terminal of the user.

Further, the information pushing device may extract intermediate sub-information from the sub-information set corresponding to each category according to the calculated information extraction number of each category, and then input the intermediate sub-information and the obtained cold start information together to the fine ranking layer to perform fine ranking processing, so as to obtain final information to be pushed, that is, the cold start information is directly transmitted to the fine ranking layer. In other embodiments of the present invention, the information pushing apparatus may further extract intermediate sub-information from each sub-information set of the recall information according to the category information to be pushed, and simultaneously extract a plurality of sub-information from the cold start information according to the category information to be pushed, for example, determine three categories of the category information to be pushed, in which the probability of user interest is the highest, and extract a plurality of sub-information from the cold start information for the three categories, and then transmit the plurality of sub-information extracted from the cold start information and the intermediate sub-information extracted from the recall information to the ranking layer for ranking to obtain final information to be pushed. Therefore, the information pushing equipment can further improve the diversity and the newness of information pushed to the user on the basis of ensuring the degree of personalized pushing for the user.

In the embodiment of the invention, the information pushing equipment can acquire the cold start information from the pushing information database, so that the information pushing equipment can screen the information to be pushed from the recall information and the cold start information according to the category information to be pushed subsequently. Therefore, the information pushing equipment can promote the diversity and the timeliness of the pushed information on the basis of ensuring the degree of personalized pushing for the user, so that the effect of personalized information pushing for the user is further promoted.

As shown in fig. 11, the information push process is divided into four main processes, namely, extracting recall information and dividing into buckets 11-1, generating an attribute feature vector 11-2, generating a category feature vector 11-3, and calculating the information extraction quantity of each category and extracting information 11-4. Specifically, the recalling information extraction and binning 11-1 is to extract initial sub-information by using a plurality of recall queues 11-11, namely recall queues 11-11a and recalls 11-11b, until recall queues 11-11n, and then perform binning and scoring 11-12 on the initial sub-information to obtain a sub-information set 11-13 corresponding to each category, wherein each sub-information in the sub-information set has a corresponding score, for example, the set 11-131: sub information 11-131 a: 0.5; sub information 11-131 b: 0.3. wherein 0.5 and 0.3 are ranking scores. When generating the attribute feature vector 11-2, the information pushing device first obtains various features 11-21 of the user, such as gender 11-211, age 11-212, region 11-213, tag 11-214, and the like, then inputs the features into the vector embedding layer 11-22 for vector conversion and splicing, and finally inputs the spliced intermediate feature vector into the attention layer 11-23 for multi-head self-attention calculation to obtain an attention parameter, and further obtain the attribute feature vector. When the category feature vector 11-3 is generated, the sub-category vector 11-31 of each historical time point, such as the vector 11-311, the vector 11-312, the vector 11-313, the vector 11-314, and the vector 11-31n, and the historical category vector output at the last time point of each historical time point, such as the vector 11-321, the vector 11-322, the vector 11-323, the vector 11-324, the vector 11-325, and the vector 11-32n, are input into the GRU unit corresponding to each historical time point, and the historical category vector at the historical time point closest to the current time, namely the vector 11-32n, is output as the category feature vector. Calculating the information extraction quantity of each category and extracting the information 11-4, namely performing vector splicing 11-41 by using attribute feature vectors and category features to obtain user feature vectors, inputting the user feature vectors into softmax11-42 to obtain category information 11-43 to be pushed, extracting intermediate sub-information 11-44 from the recall information according to the category information 11-43 to be pushed, and finally performing fine processing 11-45 on the intermediate sub-information 11-44 to obtain final information to be pushed.

Exemplarily, as shown in table 1, the data for improving the effect of the information push method provided by the embodiment of the present invention compared with the related art is as follows:

TABLE 1

As can be seen from table 1, compared with the related art, the number of click times per person is increased by 2.23%, the number of page turn over per person is increased by 3.01%, the number of exposure videos per person is increased by 2.69%, and the number of playing times per video is increased by 2.00% in the information push method provided by the embodiment of the present invention, which indicates that the user has an increased interest level in the information pushed by using the information push method provided by the embodiment of the present invention, and also indicates that the information pushed by using the information push method provided by the embodiment of the present invention is in line with the interest of the user, thereby realizing personalized information push for the user. Further, compared with the related art, the number of the second-level categories of the personally played video is increased by 3.15% and the number of the first-level categories of the personally exposed content is increased by 0.78% in the information pushing method provided by the embodiment of the invention, which indicates that the diversity of the pushed information can be increased on the premise of ensuring the personalized information pushing for the user by using the information pushing method provided by the embodiment of the invention, so that the personalized information pushing effect for the user is better.

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

The embodiment of the invention is realized in a look-at-a-look function in a social application program. Fig. 12 is a schematic diagram of information pushing in a glance according to an embodiment of the present invention. Referring to fig. 12, the process of pushing information in a visual sense can be divided into retrieving recall information from a data system 12-1 (push information database) by using an n-way recall queue 12-2, a rough ranking process 12-3 and a fine ranking process 12-4. The article set recalled by the N-way recall queue 12-2 (queue 1-N) is D ═ D₁:c₁,d₂:c₁,…,d_n:c₂In which d is_iArticle i (initial sub-information) indicating a recall, c_iIndicating the category (category label) of the article. Then, the words are classified into categories (category classification) to obtain a classified article set (initial sub-information set), for example, c₁:{d₁,d₂,…}，c₂:{d₃,d_n… } etc. Then, the characters in each bucket are scored by using a CTR model, and sorted from large to small according to scores (sorting scores) to obtain a candidate category sub-bucket set (sub-information set) C ═ C₁:{d₁:s₁,d₂:s₂,…},c₂:{d₃:s₃,d_n:s₄… }, … }, wherein s_iThe score of the ith article is represented.

The rough layout process 12-3 is divided into user portrait feature extraction 12-31, user reading history category calculation 12-32, and prediction of user reading category distribution and acquisition of in-bucket chapters 12-33. Wherein, in the user portrait feature extraction 12-31, the layer is embedded by the vector and the self-attention is paidThe mechanism layer obtains a user image feature vector (attribute feature vector). In the construction of a user image, a series of user image features (characteristic information) such as gender, age, region, label, category, etc. are used, and then a vector embedding layer is used to input U ═ U { (U) } to the user image₁,u₂,…,u_nConstructing a feature vector to obtain a user feature vector X ═ X₁,x₂,…,x_n]Then, a multi-head self-attention mechanism is used to calculate the user portrait feature vector U ═ head₁,head₂,…,head_n}W^o. Since the feature vector of the user portrait varies with time, at time t, the feature vector of the user is X_tIf the obtained user portrait feature vector is U_t. The user reading history category calculation 12 to 32 is to acquire characteristics (subcategory information) of the reading history category of the user before the current time (history time point), for example, military affairs, health preserving, marital affairs, military affairs, health preserving, and the like. After passing through the vector embedding layer, each category feature corresponds to a vector (a sub-category vector), and then the vectors are sequentially input into the GRU unit according to the time sequence to obtain a historical reading category feature vector H (a category feature vector) of the user at the time t (a historical time point closest to the current time point), so as to characterize the reading history of the user. Predicting the distribution of the user reading categories and obtaining chapters 12-33 in the barrel, namely splicing the user portrait feature vector U and the historical reading category feature vector H to obtain a final user feature vector (user feature vector), and obtaining the category distribution prediction (category information to be pushed) C ═ C of the next user through a softmax layer (trained category prediction model)₁,c₂,…,c_n}. Wherein, the calculation mode of softmax is shown as formula (6):

wherein h is_iIs vector h ═ h₁,h₂,…,h_nOne value of (c). It should be noted that softmax used in this step is trained, and labeled as a statistical userClass ratio P ═ P for x articles read₁,p₂,…,p_nThe loss function is L oss (C, P) ═ ∑_0≤i≤np·log(c_i). During training, the objective function is optimized by using a gradient descent method.

Assuming that the total number of recalled articles is T, the number of recalled articles for each category is T ═ T₁,t₂,…,t_n} (number of information extractions), where, t_i＝T×c_i. According to the method, t can be extracted from each sub-bucket of the candidate category sub-bucket set according to the principle that the score is from large to small_iAn article, as a recommendation join (intermediate sub-information).

Then, after the recommendation set continues to perform user vector and article vector representation 12-41, feature generation system 12-42 and ranking model 12-43 in the fine ranking process 12-4, a final result (information to be pushed) can be obtained, and the final result is sent to the smart phone, so that the user can see the pushed information in the column of the smart phone.

By the method, in the process of pushing at a glance, the distribution of the categories in which the user is interested in each category is predicted, and the user can know which categories are more interested in the content, so that more articles are obtained in the buckets corresponding to the categories in which the user is interested, and the obtained articles are high-quality articles with higher scores in the buckets, so that high-quality articles which may be interested in different users are pushed for the different users according to the interests of the different users, and the degree of personalized recommendation for the users is improved.

Continuing with the exemplary structure of the information pushing device 255 provided by the embodiment of the present invention implemented as software modules, in some embodiments, as shown in fig. 2, the software modules stored in the information pushing device 255 of the memory 250 may include:

the recall module 2551 is configured to, when an information push instruction is obtained, extract recall information from a push information database according to identity information of a user in the information push instruction;

an extracting module 2552, configured to extract feature information of the user according to the identity information, and acquire history category information of the user according to the identity information; the historical category information represents the category to which the information read by the user belongs;

a prediction module 2553, configured to predict category information to be pushed based on the feature information and the history category information; the category information to be pushed represents the probability that the user is interested in each category;

and the information screening module 2554 is configured to screen information to be pushed from the recall information according to the category information to be pushed, and send the information to be pushed to the terminal of the user.

In some embodiments of the present invention, the recall module 2551 is specifically configured to, when the information push instruction is acquired, extract a plurality of initial sub-information from the push information database according to the identity information; according to the category label of each piece of initial sub information, performing category division on the plurality of pieces of initial sub information to obtain an initial sub information set corresponding to each category; and utilizing all initial sub-information in the initial sub-information set to form the recall information.

In some embodiments of the present invention, the recall module 2551 is further configured to sort all initial sub information in the initial sub information set, so as to obtain a sub information set corresponding to each category;

correspondingly, the recall module 2551 is further configured to utilize all sorted initial sub-information in the sub-information set to compose the recall information.

In some embodiments of the present invention, the recall module 2551 is further specifically configured to obtain a content feature of the initial sub-information, and calculate a ranking score of the initial sub-information by using the content feature; and sequencing the initial sub-information by using the sequencing fraction to obtain sequenced initial sub-information, and forming the sub-information set by using the sequenced initial sub-information.

In some embodiments of the present invention, the prediction module 2553 is specifically configured to construct an attribute feature vector of the user by using the feature information; determining a category characteristic vector of the user according to the historical category information; wherein the category feature vector can characterize the characteristics of the historical category information; and predicting the category information to be pushed by utilizing the attribute feature vector and the category feature vector.

In some embodiments of the present invention, the prediction module 2553 is specifically configured to perform vector transformation on each of the one or more sub-feature information to obtain one or more sub-feature vectors; splicing an intermediate feature vector by using the one or more sub-feature vectors; performing self-attention calculation on the intermediate feature vector for multiple times to obtain multiple attention parameters corresponding to the intermediate feature vector; wherein the attention parameter characterizes the importance degree of the sub-feature information; and splicing the attention parameters to obtain the attribute feature vector.

In some embodiments of the present invention, the prediction module 2553 is specifically configured to perform vector transformation on the sub-category information to obtain a sub-category vector; predicting historical category vectors corresponding to each historical time point one by utilizing the corresponding subcategory vector of each historical time point and the historical category vector of the last time point of each historical time point; and taking the history category vector corresponding to the history time point closest to the current time point as the category feature vector.

In some embodiments of the present invention, the prediction module 2553 is specifically configured to splice the attribute feature vectors and the category feature vectors to obtain user feature vectors; and predicting the category information of the user characteristic vector by using the trained category prediction model to obtain the category information to be pushed.

In some embodiments of the present invention, the information screening module 2554 is specifically configured to calculate, according to the category information to be pushed, an information extraction number for each sub information set in the recall information; extracting sub-information from each sub-information set according to the information extraction quantity to obtain intermediate sub-information; and performing fine arrangement processing on the intermediate sub-information to obtain the information to be pushed.

In some embodiments of the present invention, the information filtering module 2554 is specifically configured to start obtaining sub information from the sub information with the highest ranking score in each sub information set until obtaining the information extraction quantity sub information; and taking the obtained sub information as the middle sub information.

In some embodiments of the present invention, the recall module 2551 is further configured to obtain cold start information from the push information database; the cold start information refers to information that the read times are lower than a preset time threshold;

correspondingly, the information screening module 2554 is further configured to screen information to be pushed from the recall information and the cold start information according to the category information to be pushed, and send the information to be pushed to the terminal of the user.

Embodiments of the present invention provide a computer-readable storage medium storing executable instructions, where the executable instructions are stored, and when executed by a processor, the executable instructions cause the processor to execute an information pushing method provided by embodiments of the present invention, for example, the method shown in fig. 3, fig. 5, fig. 6 and fig. 10.

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

In some embodiments, executable instructions may be written in any form of programming language (including compiled or interpreted languages), in the form of programs, software modules, scripts or code, and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

By way of example, executable instructions may correspond, but do not necessarily correspond, to files in a file system, may be stored in a portion of a file that holds other programs or data, such as in one or more scripts stored in a hypertext markup language (HTM L, HyperTextMarkup L engine) 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).

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

The above description is only an example of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, and improvement made within the spirit and scope of the present invention are included in the protection scope of the present invention.

Claims

1. An information pushing method, comprising:

2. The method according to claim 1, wherein when the information push instruction is obtained, extracting recall information from a push information database according to the identity information of the user in the information push instruction comprises:

when the information pushing instruction is obtained, extracting a plurality of initial sub-information from the pushing information database according to the identity information;

according to the category label of each piece of initial sub information, performing category division on the plurality of pieces of initial sub information to obtain an initial sub information set corresponding to each category;

and utilizing all initial sub-information in the initial sub-information set to form the recall information.

3. The method according to claim 2, wherein after the performing category classification on the plurality of initial sub-information according to the category label corresponding to each initial sub-information to obtain an initial sub-information set corresponding to each category, before the composing the recall information by using all initial sub-information in the initial sub-information set, the method further comprises:

sequencing all initial sub information in the initial sub information set to obtain a sub information set corresponding to each category;

correspondingly, the composing the recall information by using all initial sub-information in the initial sub-information set includes:

and utilizing all the sorted initial sub-information in the sub-information set to form the recall information.

4. The method of claim 3, wherein the sorting the initial sub information in the initial sub information set to obtain the sub information set corresponding to each category comprises:

acquiring content characteristics of the initial sub-information in the initial sub-information set, and calculating the ranking score of the initial sub-information by using the content characteristics;

and sequencing the initial sub-information by using the sequencing fraction to obtain sequenced initial sub-information, and forming the sub-information set by using the sequenced initial sub-information.

5. The method according to any one of claims 1 to 4, wherein predicting the category information to be pushed based on the feature information and the historical category information comprises:

constructing an attribute feature vector of the user by using the feature information;

determining a category characteristic vector of the user according to the historical category information; wherein the category feature vector can characterize the characteristics of the historical category information;

and predicting the category information to be pushed by utilizing the attribute feature vector and the category feature vector.

6. The method of claim 5, wherein the feature information comprises one or more sub-feature information; the constructing an attribute feature vector of the user by using the feature information includes:

performing vector transformation on each piece of sub-feature information in the one or more pieces of sub-feature information to obtain one or more sub-feature vectors;

splicing an intermediate feature vector by using the one or more sub-feature vectors;

performing self-attention calculation on the intermediate feature vector for multiple times to obtain multiple attention parameters corresponding to the intermediate feature vector; wherein the attention parameter characterizes the importance degree of the sub-feature information;

and splicing the attention parameters to obtain the attribute feature vector.

7. The method of claim 5, wherein the historical category information includes subcategory information corresponding to each historical point in time; the determining the category feature vector of the user according to the historical category information includes:

performing vector conversion on the subcategory information to obtain a subcategory vector;

predicting historical category vectors corresponding to each historical time point one by utilizing the corresponding subcategory vector of each historical time point and the historical category vector of the last time point of each historical time point;

and taking the history category vector corresponding to the history time point closest to the current time point as the category feature vector.

8. The method according to claim 5, wherein the predicting the category information to be pushed by using the attribute feature vector and the category feature vector comprises:

splicing the attribute feature vector and the category feature vector to obtain a user feature vector;

and predicting the category information of the user characteristic vector by using the trained category prediction model to obtain the category information to be pushed.

9. The method according to any one of claims 1 to 4 or 6 to 8, wherein the screening information to be pushed from the recall information according to the category information to be pushed comprises:

calculating the information extraction quantity aiming at each sub information set in the recall information according to the category information to be pushed;

extracting sub-information from each sub-information set according to the information extraction quantity to obtain intermediate sub-information;

and performing fine arrangement processing on the intermediate sub-information to obtain the information to be pushed.

10. The method according to claim 9, wherein said extracting sub information from each sub information set according to the information extraction number to obtain intermediate sub information comprises:

starting the acquisition of sub information from the sub information with the highest sorting score in each sub information set until the information extraction quantity sub information is acquired;

and taking the obtained sub information as the middle sub information.

11. The method according to any one of claims 1 to 4, 6 to 8, or 10, wherein before the filtering out information to be pushed from the recalled information according to the category information to be pushed and sending the information to be pushed to the terminal of the user, the method further comprises:

acquiring cold start information from the push information database; the cold start information refers to information that the read times are lower than a preset time threshold;

correspondingly, the screening out information to be pushed from the recall information according to the category information to be pushed and sending the information to be pushed to the terminal of the user includes:

and screening the information to be pushed from the recall information and the cold start information according to the category information to be pushed, and sending the information to be pushed to the terminal of the user.

12. An information pushing apparatus, comprising:

13. An information push apparatus characterized by comprising:

the memory is used for storing an executable information pushing instruction;

a processor configured to implement the method of any one of claims 1 to 11 when executing the executable information push instructions stored in the memory.

14. A computer-readable storage medium having stored thereon executable information push instructions for causing a processor to perform the method of any one of claims 1 to 11 when executed.