CN110781391A - Information recommendation method, device, equipment and storage medium - Google Patents

Information recommendation method, device, equipment and storage medium Download PDF

Info

Publication number
CN110781391A
CN110781391A CN201911005267.9A CN201911005267A CN110781391A CN 110781391 A CN110781391 A CN 110781391A CN 201911005267 A CN201911005267 A CN 201911005267A CN 110781391 A CN110781391 A CN 110781391A
Authority
CN
China
Prior art keywords
recommendation
information
rate
model
feature
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201911005267.9A
Other languages
Chinese (zh)
Other versions
CN110781391B (en
Inventor
唐红艳
赵铭
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen Yayue Technology Co ltd
Original Assignee
Tencent Technology Shenzhen Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tencent Technology Shenzhen Co Ltd filed Critical Tencent Technology Shenzhen Co Ltd
Priority to CN201911005267.9A priority Critical patent/CN110781391B/en
Publication of CN110781391A publication Critical patent/CN110781391A/en
Application granted granted Critical
Publication of CN110781391B publication Critical patent/CN110781391B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/954—Navigation, e.g. using categorised browsing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/958—Organisation or management of web site content, e.g. publishing, maintaining pages or automatic linking

Landscapes

  • Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Remote Sensing (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention provides an information recommendation method, device, equipment and storage medium; the method comprises the following steps: receiving an information browsing request of a target object, and acquiring a historical browsing record of the target object according to the information browsing request; obtaining a recommendation information set aiming at a target object according to a historical browsing record; obtaining recommendation characteristics corresponding to each recommendation information from each recommendation information in the recommendation information set and historical browsing records; the recommendation characteristics represent the combined information of each piece of recommendation information and the historical browsing records; calculating recommendation parameter information for each recommendation information by using a preset recommendation model and the recommendation characteristics; the recommendation parameter information comprises at least one of playing completion degree, effective playing rate, positive feedback rate and screen interaction rate; and recommending the recommendation information set based on the recommendation parameter information. By the method and the device, the accuracy of information recommendation can be improved.

Description

Information recommendation method, device, equipment and storage medium
Technical Field
The present invention relates to recommendation system technologies, and in particular, to an information recommendation method, apparatus, device, and storage medium.
Background
The information recommendation refers to a process of recommending contents which may be interested by a user for the user according to historical browsing records of the user on various types of information. In an information recommendation system, generally, an information recommendation model is used to analyze various operations of a user on information in a history browsing record, find out information contents that the user may be interested in from an information base, sort the information, and display the information on a terminal interface of the user.
Currently, a commonly used information recommendation method is to predict an effective play rate of information, so as to sort recommendation information according to the effective play rate, however, information recommendation is performed only by using the effective play rate, so that the information recommendation dimension is single, and thus the accuracy rate of information recommendation performed for a user is low.
Disclosure of Invention
The embodiment of the invention provides an information recommendation method, an information recommendation device and a storage medium, which can improve the accuracy of information recommendation.
The technical scheme of the embodiment of the invention is realized as follows:
in a first aspect, an embodiment of the present invention provides an information recommendation method, including:
receiving an information browsing request of a target object, and acquiring a historical browsing record of the target object according to the information browsing request;
obtaining a recommendation information set aiming at the target object according to the historical browsing record; the recommendation information set is a set consisting of recommendation information;
obtaining recommendation characteristics corresponding to each recommendation information from each recommendation information in the recommendation information set and the historical browsing records; the recommendation characteristics represent the combined information of each recommendation information and the historical browsing records;
calculating recommendation parameter information for each piece of recommendation information by using a preset recommendation model and the recommendation characteristics; the recommendation parameter information comprises at least one of playing completion degree, effective playing rate, positive feedback rate and screen interaction rate; the preset recommendation model is a model established by taking the recommendation parameter information as a target;
and recommending the recommendation information set based on the recommendation parameter information.
In a second aspect, an embodiment of the present invention provides an information recommendation apparatus, including:
the receiving module is used for receiving an information browsing request of a target object and acquiring a historical browsing record of the target object according to the information browsing request;
the characteristic extraction module is used for obtaining a recommendation information set aiming at the target object according to the historical browsing record of the target object; the recommendation information set is a set consisting of recommendation information; obtaining recommendation characteristics corresponding to each recommendation information from each recommendation information in the recommendation information set and the historical browsing records; the recommendation characteristics represent the combined information of each recommendation information and the historical browsing records;
the recommendation module is used for calculating recommendation parameter information aiming at each piece of recommendation information by utilizing a preset recommendation model and the recommendation characteristics; the recommendation parameter information comprises at least one of a playing completion degree, an effective playing rate, a positive feedback rate and the screen interaction rate; the preset recommendation model is a model established by taking the recommendation parameter information as a target; and recommending the recommendation information set based on the recommendation parameter information.
In a third aspect, an embodiment of the present invention provides an information recommendation apparatus, including:
a memory for storing executable information recommendation instructions;
a processor configured to implement the method according to the first aspect when executing the executable information recommendation instruction stored in the memory.
In a fourth aspect, an embodiment of the present invention provides a storage medium, where executable information recommendation instructions are stored, and the executable information recommendation instructions are configured to cause a processor to implement the method according to the first aspect.
The embodiment of the invention has the following beneficial effects:
the information recommending device determines a recommended information set according to the historical browsing records of the target object, then determines the combined information of each piece of recommended information and the historical browsing records according to each piece of recommended information in the recommended information set and the historical browsing records of the target object, and then determines recommended parameter information according to the combined information, wherein the recommended parameter information can comprise at least one of playing completion, effective playing rate, positive feedback rate and screen interaction rate, and then recommends the recommended information set according to the recommended parameter information.
Drawings
Fig. 1 is an alternative structural diagram of an information recommendation system according to an embodiment of the present invention;
fig. 2 is a schematic structural diagram of an information recommendation device according to an embodiment of the present invention;
fig. 3 is a schematic diagram illustrating an alternative flow of an information recommendation method according to an embodiment of the present invention;
fig. 4 is a schematic interface diagram of a terminal sending an information browsing request according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of a recommendation provided by an embodiment of the present invention;
fig. 6 is a schematic diagram illustrating an optional flow of an information recommendation method according to an embodiment of the present invention;
FIG. 7 is a schematic diagram of a get recommendations feature provided by embodiments of the present invention;
FIG. 8 is a diagram of a sharing layer according to an embodiment of the present invention;
FIG. 9 is a diagram of a pre-set recommendation model according to an embodiment of the present invention;
FIG. 10 is a diagram illustrating a statistical common category label according to an embodiment of the present invention;
fig. 11 is a schematic diagram illustrating an alternative flow chart of an information recommendation method according to an embodiment of the present invention;
FIG. 12 is a schematic diagram of a video recommendation model in a short video recommendation scenario provided by an embodiment of the present invention;
fig. 13 is a schematic diagram of a short video recommendation process provided by 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 following description, references to the terms "first \ second \ third \ fourth" are only to distinguish similar objects and do not denote a particular order or importance to the objects, and it is to be understood that "first \ second \ third \ fourth" may be interchanged with a particular order or sequence as appropriate to enable the embodiments of the invention described herein to be practiced in an order other than that shown 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) The information recommendation system is characterized by judging the interest of a user by analyzing various historical operations of the user, such as playing, praise, comment and the like, so that information which may be interested in the user is recommended for the user.
2) And the information browsing request is used for informing the information recommendation system of a request for refreshing the presented information when representing that the user needs to browse the information. For example, a user opening an application program or clicking a refresh button in the application program may be regarded as sending an information browsing request, at this time, the information recommendation system refreshes the presented information, and then presents the information obtained by the refreshing on a display interface of the terminal of the user.
3) And recommending an information set, wherein the information set represents the set of information which needs to be presented when the device is refreshed. When presenting information, the information recommendation system needs to sort the information in the recommendation information set, so that the information in the recommendation information set is presented in order.
4) The Play Completion Rate (PCR) represents the Play progress of the information presented by the information recommendation system predicted by the information recommendation system, i.e. the length of the user's browsing, and is proportional to the total length of the presented information. For example, when the information presented by the information recommendation system is a video, the playing completion degree refers to the part of the video that has already been played, and occupies the proportion of the total length of the video.
5) The effective playing Rate (CTR) is a probability that the presented information predicted by the information recommendation system is effectively played, i.e. a probability that the user effectively browses the information presented by the information recommendation system. For example, for a certain video presented by the information recommendation system, when it is specified that the user watches more than 60% of the content, the effective playing is considered as effective playing, and in this case, the effective playing rate refers to the probability that the user watches more than 60% of the content of the video predicted by the information recommendation system.
6) The Positive Feedback Rate (PFR) refers to a probability that a user has Positive Feedback on information presented by the information recommendation system, which is predicted by the information recommendation system, and the Positive Feedback may be operations such as like and collection, which may represent a propagation will and an interest level of the user on the information.
7) The Screen Interaction Rate (SIR) represents the probability of Screen Interaction performed by the user on the information presented by the information recommendation system, which is predicted by the information recommendation system. The screen interaction may refer to an operation of controlling information presented by the recommendation information through an operation on the display interface by the user, for example, operations of full screen, back, forward, and zoom-in, which may all indicate that the user has a high interest in the information.
8) And comment information, which represents the probability of commenting the information presented by the information recommendation system by the user predicted by the information recommendation system. When the user comments on the information presented by the information recommendation system, the fact that the user pays more attention to the information presented by the information recommendation system is indicated.
9) The recommendation score represents the interest degree of the user in the recommendation information calculated by the information recommendation system, and can be used as a basis for the information recommendation system to sort the recommendation information in the recommendation information set.
The information recommendation refers to a process of analyzing various operations of the information in the historical browsing records by the user, finding out the content which is possibly interested by the user from the information base, and orderly presenting the content which is possibly interested by the user on a display interface of the terminal of the user.
In the related art, information recommendation may be performed for a user by an effective play rate, information recommendation may be performed for a user by a plurality of independent models, information recommendation may be performed for a user by a multitask model, and information recommendation may be performed for a user by a fixed weight multitask model.
The information recommendation is performed on the user information through the effective playing rate, namely, the information recommendation system predicts the probability of the user effectively playing the information in the information base, and then sorts the information in the information base according to the predicted probability to obtain the recommendation information. However, the method only considers the probability of playing the information by the user, omits operations such as positive feedback of the user to the information, screen interaction and the like which can represent the interest degree of the user to the information, and has single recommendation dimension, so that the information recommendation accuracy is low.
The information recommendation method includes the steps that a plurality of independent models are used for information recommendation for users, namely, the information recommendation system respectively establishes models for users aiming at operations such as playing, praise and comment of information, then probabilities of the users respectively conducting operations such as playing, praise and comment on the information in an information base are respectively predicted by the aid of the established models, finally the probabilities of the operations are combined, and the information in the information base is sorted by the aid of the combined probabilities. However, in this method, modeling is performed separately with a single operation as a target, and there is a high possibility that there is a deviation between the training samples and the samples to be predicted, for example, when modeling is performed with the playing completion degree as a target, the training samples are playing samples, and prediction is performed on the entire samples at the time of prediction, so that the prediction accuracy is low. Even if information recommendation is performed by using multiple recommendation dimensions, the accuracy of information recommendation is low because the prediction result of a sample needing prediction by using multiple independent models is not ideal.
The method is characterized in that a multitask learning method is utilized to model playing, praise, comment and other operations as targets, during training, an embedded layer and a part of all-connected layers at the bottom layer share parameters, and independent layers corresponding to all targets at the upper layer are only affected by the targets. When information is recommended, a multitask model is used for predicting various operations of the user aiming at the information, and then information recommendation is performed on the user by combining prediction results of the various operations. However, due to the unbalanced label distribution of the training samples in different operations, the multi-task model learning is insufficient, for example, in the case of screen interaction, the screen interaction behavior of the user is sparse in the whole training sample, so that the amount of negative samples of screen interaction in the training sample is far greater than the amount of positive samples, the screen interaction target learning is insufficient, and the prediction accuracy of response screen interaction is high. Therefore, due to the fact that label distribution of the training samples is not balanced, prediction results of some operations are not accurate, and accuracy of information recommendation by using the multi-task model is low.
The information recommendation is carried out by utilizing the fixed weight multitask model, namely the loss functions of different targets are weighted on the basis of the information recommendation carried out by the multitask model, and the integral loss function of the fixed weight multitask model is obtained. During prediction, the probability of each operation of the user is still predicted by using the multitask model, and information recommendation is further realized. However, since the weights of the penalty function of the fixed-weight multitask model are hyper-parameters, they need to be set before training starts and they do not change once training starts. When the level or the dependency relationship exists among the targets, different weights need to be used in different training states, so that the fixed weight multitask model cannot achieve the optimal training effect, accurate prediction results cannot be obtained for various operations, and the accuracy of information recommendation cannot be improved even if information recommendation is performed by using multiple recommendation dimensions.
As can be seen from the above, in the related art, when information recommendation is performed only by using the effective play rate, the accuracy of information recommendation is low because of a single recommendation dimension, and when information recommendation is performed by using a plurality of independent task models, a multitask model and a fixed weight multitask model, the accuracy of information recommendation is not effectively improved.
The embodiment of the invention provides an information recommendation method, an information recommendation device, information recommendation equipment and a storage medium, which can improve the accuracy of information recommendation. An exemplary application of the information recommendation device provided by the embodiment of the present invention is described below, and the information recommendation device provided by 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 recommendation apparatus is implemented as a server will be described.
Referring to fig. 1, fig. 1 is an alternative architecture diagram of an information recommendation system 100 according to an embodiment of the present invention, in order to implement supporting an information recommendation application, a terminal 400 (an exemplary terminal 400-1 and a graphical interface 410-1 of the terminal 400-1) 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. The server 200 is configured with a database 500, and the database 500 stores the browsing history of the user and all the information in the information recommendation system 100.
After the user wakes up the terminal 400-1, the user can enter a pre-program of the information recommendation system 100 on the terminal 400-1 through the graphical interface 410-1 of the terminal 400-1, and by clicking a refresh button, the user is indicated to the information recommendation system 100 that the user needs to browse information, and the information recommendation system 100 is requested to perform information recommendation for the user, at this time, the terminal 400-1 sends an information browsing request to the server 200 through the network 300. The server 200 receives an information browsing request sent by the terminal 400-1, acquires a historical browsing record of the user from the database 500 according to the information browsing request, analyzes the information category interested by the user according to the historical browsing record of the user, and determines a set of information possibly interested by the user from the database 500 according to the information category interested by the user, that is, determines a recommended information set. Next, the server 200 needs to determine an order for each recommendation information in the recommendation information set. The server 200 combines each piece of recommendation information in the recommendation information set with the historical browsing records to obtain combined information of each piece of recommendation information and the historical browsing records, and extracts recommendation features from the combined information. Then, the server 200 inputs the recommendation characteristics into a trained preset recommendation model for operation, predicts recommendation parameter information of each recommendation information respectively, and then determines the interest degree of the target object in each recommendation information according to the predicted recommendation parameter information, i.e. calculates a recommendation score for each recommendation information, wherein the recommendation parameter information includes at least one of a playing completion degree, an effective playing rate, a positive feedback rate and a screen interaction rate. Then, the server 200 sorts each piece of information recommendation information in the recommendation information set according to the sequence of recommendation scores from high to low, and transmits the final sorting result to the terminal 400-1 through the network 200, and the terminal 400-1 displays the sorting result on the graphical interface 410-1 so as to present the sorting result to the server 200, thereby completing the information recommendation process.
Referring to fig. 2, fig. 2 is a schematic structural diagram of an information recommendation apparatus 600 according to an embodiment of the present invention, and the information recommendation apparatus 600 shown in fig. 2 includes: at least one processor 610, memory 650, at least one network interface 620, and a user interface 630. The various components in the information recommendation device 600 are coupled together by a bus system 640. It is understood that bus system 640 is used to enable communications among the components. Bus system 640 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 640 in fig. 2.
The Processor 610 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 630 includes one or more output devices 631 including one or more speakers and/or one or more visual displays that enable the presentation of media content. The user interface 630 also includes one or more input devices 632, including user interface components to facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.
The memory 650 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 depicted memory 650 of embodiments of the invention is intended to comprise any suitable type of memory. Memory 650 optionally includes one or more storage devices physically located remote from processor 610.
In some embodiments, memory 650 can store data to support various operations, examples of which include programs, modules, and data structures, or subsets or supersets thereof, as exemplified below.
An operating system 651 including system programs for handling various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and for handling hardware-based tasks;
a network communication module 652 for reaching other computing devices via one or more (wired or wireless) network interfaces 620, exemplary network interfaces 620 including: bluetooth, wireless compatibility authentication (WiFi), and Universal Serial Bus (USB), etc.;
a display module 653 for enabling the presentation of information (e.g., a user interface for operating peripherals and displaying content and information) via one or more output devices 631 (e.g., a display screen, speakers, etc.) associated with the user interface 630;
an input processing module 654 for detecting one or more user inputs or interactions from one of the one or more input devices 632 and translating the detected inputs or interactions.
In some embodiments, the information recommendation apparatus provided by the embodiments of the present invention may be implemented in software, and fig. 2 shows the information recommendation apparatus 655 stored in the memory 650, which may be software in the form of programs and plug-ins, and includes the following software modules: a receiving module 6551, a feature extraction module 6552 and a recommendation module 6553, the functions of each of which will be described below.
In other embodiments, the information recommendation apparatus provided in embodiments of the present invention may be implemented in hardware, and for example, the information recommendation apparatus provided in embodiments of the present invention may be a processor in the form of a hardware decoding processor, which is programmed to execute the information recommendation method provided in embodiments of the present invention, for example, the processor in the form of the hardware decoding processor may be one or more Application Specific Integrated Circuits (ASICs), DSPs, Programmable Logic Devices (PLDs), Complex Programmable Logic Devices (CPLDs), Field Programmable Gate Arrays (FPGAs), or other electronic components.
Illustratively, an embodiment of the present invention provides an information recommendation apparatus, including:
a memory for storing executable information recommendation instructions;
and the processor is used for realizing the information recommendation method provided by the embodiment of the invention when the executable information recommendation instruction stored in the memory is executed.
The information recommendation method provided by the embodiment of the invention will be described in conjunction with the exemplary application and implementation of the information recommendation device provided by the embodiment of the invention.
Referring to fig. 3, fig. 3 is an optional flowchart illustration of an information recommendation method according to an embodiment of the present invention, where the information recommendation method according to the embodiment of the present invention may include:
s101, receiving an information browsing request of a target object, and acquiring a historical browsing record of the target object according to the information browsing request.
The embodiment of the invention is realized in the scene that the target object needs to browse the information. The target object wakes up the terminal and sends an information browsing request to the information recommendation device through the terminal, the information recommendation device receives the information browsing request of the target object, and then the historical browsing record of the target object is obtained from the database according to the information browsing request.
It should be noted that, in the embodiment of the present invention, the information browsing request may be issued when the target object opens the information recommendation application on the terminal, or may be issued when the target object performs a refresh operation on the information recommendation application, which is not specifically limited herein.
It is to be understood that, in the embodiment of the present invention, the target object refers to any user that has been registered in the information recommendation application.
In the embodiment of the invention, the information browsing request carries the identity of the target object so as to indicate to the information recommendation device which user the information browsing request needs to recommend information. The identity of the target object may be account information of the user, a nickname of the user, or any other identity that can uniquely indicate the identity of the user, and the identity of the specific target object may be set according to actual requirements.
Illustratively, the embodiment of the present invention provides an interface schematic for a terminal to send an information browsing request, as shown in fig. 4, in a display area 1 of a display interface of the terminal, a search box and a refresh button are provided. In the display area 2, three tabs of attention, recommendation and hot list are displayed, and information recommendation can be performed by the information recommendation device for a target object according to the tabs. In the display area 3, the original recommendation information presented by the information recommendation apparatus through the terminal is displayed, which is title 1, content abstract 1, title 2, content abstract 2, title 3, and content abstract 3. The target object clicks a refresh button in the display area 1, and sends an information browsing request to the information recommendation device, wherein the information browsing request contains account information of the target object, so that the subsequent information recommendation device carries out information recommendation on the target object according to the information browsing request.
S102, obtaining a recommendation information set aiming at a target object according to the historical browsing record; the recommendation information set is a set composed of recommendation information.
After acquiring the historical browsing records of the target object from the database, the information recommendation device may analyze the historical browsing records of the target object to obtain recommendation information to be recommended to the target object, and use the recommendation information to form a recommendation information set.
In this embodiment of the present invention, the recommendation information may be an information set composed of all information stored in the database, or may be a set composed of a part of information stored in the database, and the embodiment of the present invention is not limited herein.
In the embodiment of the present invention, the recommendation information may include identification information of the information, such as an information ID, a title, a cover page, and the like, and may also include other attribute information of the information, such as a category, a label, and the like.
It is understood that the recommendation information may be information in text form, such as news, popular articles, etc., information in picture form, such as photos, moving images, etc., and information in video form, such as short videos, etc. Of course, the recommendation information may also be information in other forms, and the embodiment of the present invention is not limited in detail herein.
In some embodiments of the present invention, the information recommendation apparatus may analyze the historical browsing history of the target object to obtain information categories that may be of interest to the target object, and then use each piece of information belonging to the information category in the database as the recommendation information and use the recommendation information as the recommendation information set.
It should be noted that the history browsing record includes an identifier, a category, a tag, and the like of each piece of information browsed by the target object before the target object starts to recommend the piece of information, for example, an information title, an information category, and an information ID; the operation and interaction of the target object on each piece of browsed information can be further included, for example, the target object approves and forwards the information, and page enlargement, screen scrolling, fast forwarding and the like are performed by the target object when the information is browsed. The content included in the specific history browsing record may be set according to actual requirements, and the embodiment of the present invention is not limited herein.
S103, obtaining recommendation characteristics corresponding to each recommendation information from each recommendation information in the recommendation information set and historical browsing records; the recommendation characteristics represent the combination information of each recommendation information and the historical browsing records.
The information recommending device extracts the features of each piece of recommended information and each historical browsing record in the recommended information set to obtain the feature of each piece of recommended information and the feature of each historical browsing record, and then combines the features of the historical browsing records and the features of each piece of recommended information to obtain the recommended feature corresponding to each piece of recommended information.
It is understood that, since the recommendation feature corresponds to each recommendation information, the number of recommendation features obtained by the information recommendation device is the same as the number of recommendation information in the recommendation information set.
It should be noted that the recommended features include features corresponding to all information in the history browsing record and features corresponding to each piece of recommended information. For example, when the history browsing record includes the identifier, category, tag, etc. of the information browsed by the target object, and the operation, interaction, etc. performed by the target object for the browsed information, the recommended features include features corresponding to the identifier, the category, the tag, the operation, and the interaction of the information browsed by the target object; the recommendation information includes an information ID, a title, and a category to which the information belongs, and in this case, the recommendation feature further includes a feature corresponding to the information ID of the recommendation information, a feature corresponding to the title, and a feature corresponding to the category to which the information belongs.
In the embodiment of the invention, the recommended features need to be input into the preset recommendation model for calculation in the subsequent steps, so that the form of the recommended features needs to be the same as the input form of the preset recommendation model.
Illustratively, when the input of the preset recommendation model is a vector, the recommendation features need to be characterized in a vector form, and when the input of the preset recommendation model is a matrix, the recommendation features need to be correspondingly characterized in a matrix form.
It can be understood that the process of obtaining the recommendation characteristic corresponding to each piece of recommendation information from each piece of recommendation information in the recommendation information set and the historical browsing records can be regarded as converting the historical browsing records and the content contained in each piece of recommendation information into the same form as the input of the preset recommendation model. For example, when the input of the preset recommendation model is a vector, the information recommendation device converts the content contained in the history browsing history and each piece of recommendation information into a vector form, and when the input of the preset recommendation model is a matrix, the information recommendation device converts the content contained in the history browsing history and each piece of recommendation information into a matrix form.
S104, calculating recommendation parameter information for each piece of recommendation information by using a preset recommendation model and recommendation characteristics; the recommendation parameter information comprises at least one of playing completion degree, effective playing rate, positive feedback rate and screen interaction rate; the preset recommendation model is a model established by taking recommendation parameter information as a target.
After obtaining the recommendation characteristics corresponding to each piece of recommendation information, the information recommendation device first obtains a pre-stored preset recommendation model from a storage space of the information recommendation device, and then inputs the recommendation characteristics into the preset recommendation model for calculation to obtain the recommendation parameter information predicted by the preset recommendation model. The recommendation parameter information includes at least one of a playing progress, namely a playing completion degree, of each recommendation information, a probability that each recommendation information is effectively played, namely an effective playing rate, a probability that each recommendation information is positively fed back when being played, namely a positive feedback rate, and a probability that each recommendation information is interacted when being played, namely a screen interaction rate.
It can be understood that the preset recommendation model is a model established before information recommendation is performed on the target object, and the preset recommendation model is a model established with recommendation parameter information as a target, that is, the preset recommendation model is a model established with at least one of the playing completion degree, the effective playing rate, the positive feedback rate and the screen interaction rate as a target.
It should be noted that, in the embodiment of the present invention, since the playing completion is a proportion of a portion browsed by the target object to the total length of the recommendation information, a value range of the playing completion is [0, 1 ]; the effective play rate is the probability that the recommended information predicted by the preset recommendation model is effectively played, so the value range of the effective play rate is also between [0 and 1 ]; similarly, the positive feedback rate refers to the probability of positive feedback of the recommendation information predicted by the preset recommendation model, and the screen interaction rate refers to the probability of screen interaction of the target object predicted by the preset recommendation model in the recommendation information presentation process, so the value ranges of the positive feedback rate and the screen interaction rate are [0, 1 ].
For example, in the embodiment of the present invention, when the recommendation information is video-type information, the playing completion degree may be a time length of a video that has been played, which is a proportion of a total time length of the video; when the recommended information is character information, the playing completion degree can be the number of words already displayed on the display interface of the terminal and accounts for the proportion of the total number of words of the article; when the recommendation information is image information, the playing completion degree may be the number of image frames already displayed on the display interface of the terminal, which is a proportion of the total number of image frames.
In some embodiments of the present invention, the effective play may be related to the play completion and the play duration of the recommendation information. For example, the effective playing may be defined as a playing operation in which the playing time length exceeds a preset time, may be defined as an operation in which the playing completion degree exceeds a preset threshold, and may be defined as an operation in which the playing time length exceeds a preset time and the playing completion degree exceeds a preset time. The specific effective playing may be set according to actual conditions, and the embodiment of the present invention is not limited in particular herein.
For example, when the recommendation information is a short video with a duration of less than 15s, a play operation with a play duration of more than 5s may be regarded as an effective play, and when the recommendation information is a video with a duration of more than 15s, a play operation with a play completion degree of more than 50% may be regarded as an effective play; when the recommendation information is an article, an operation in which the playback completion degree exceeds 40% may be regarded as effective playback.
It is understood that, in the embodiment of the present invention, positive feedback refers to any operation that can characterize the target object as being interested in recommendation information.
For example, the positive feedback may be an operation such as praise and the like which can represent that the target object has an evaluation intention on the recommendation information, and may also be an operation such as forwarding and sharing which represents that the target object has a propagation intention on the recommendation information, and these operations all indicate that the target object has a high interest on the recommendation information.
It should be noted that the screen interaction means control operations of the target object on the recommendation information during the recommendation information playing process, and these control operations may reflect that the target object has a higher viewing interest on the recommendation information being played. For example, when the recommendation information is a video, the screen interaction may refer to operations of a target object such as full screen, horizontal screen, fast forward, and backward during the playing of the recommendation information, and when the recommendation information is an article, the screen interaction may refer to operations of a target object such as zooming in and zooming out on a page during the playing of the recommendation information.
In some embodiments of the present invention, the information recommendation device can not only directly obtain the playing completion, the effective playing rate, the positive feedback rate and the screen interaction rate through the prediction of the preset recommendation model, so as to calculate the recommendation score according to at least one of the playing completion degree, the effective playing rate, the positive feedback rate and the screen interaction rate, and predict the playing completion degree, the effective playing rate, the positive feedback rate and the screen interaction rate through a preset recommendation model, then using the product result of the effective playing rate and the playing completion degree as the final playing completion degree, using the product result of the effective playing rate and the screen interaction rate as the final screen interaction rate, and then calculating a recommendation score according to at least two of the final playing completion, the effective playing rate, the positive feedback rate and the final screen interaction rate, which is not limited herein.
It should be noted that the preset recommendation model in the embodiment of the present invention may be a Deep learning model, such as a Deep Neural Network (DNN), or may be a shallow Machine learning model, such as a Support Vector Machine (SVM). The specific preset recommendation model may be selected according to actual requirements, and the embodiment of the present invention is not limited herein.
And S105, recommending the recommendation information set based on the recommendation parameter information.
After the information recommending device obtains the recommendation parameter information, the information recommending device judges the interest degree of the target object for each recommendation information according to the recommendation parameter information, recommends the recommendation information set according to the interest degree of the target object for each recommendation information, returns the recommendation result to the terminal of the target object, and presents the recommendation result to the target object through the terminal, so that the information recommending device completes the information recommending process.
For example, as shown in fig. 5, the information recommendation device presents the recommendation result to the target object through the terminal, and a recommendation page is displayed in a display interface of the terminal, where the recommendation information is arranged in a descending order according to the interest degree of the target object for each piece of recommendation information. The information recommendation device also displays the interest degree of the target object for each recommendation information in a recommendation page in a recommendation degree mode. Referring to fig. 5, on the recommendation page, recommendation information 1, an abstract 1 of the recommendation information 1, and a recommendation degree 95% of the recommendation information 1 are displayed; recommendation information 2, abstract 2 of recommendation information 2, and recommendation degree of recommendation information 2 is 80%; the recommendation information 3, the abstract 3 of the recommendation information 3, and the recommendation degree of the recommendation information 3 are 70%, so that the information recommendation device finishes presenting the recommendation result.
In the embodiment of the invention, the information recommendation device can determine a recommendation information set according to the historical browsing record of the target object, then determine the combined information of each recommendation information and the historical browsing record according to each recommendation information in the recommendation information set and the historical browsing record of the target object, then determine recommendation parameter information according to the combined information, wherein the recommendation parameters can comprise at least one of playing completion degree, effective playing rate, front feedback rate and screen interaction rate, and then recommend the recommendation information set according to the recommendation parameter information, so that the information recommendation device can evaluate the recommendation information by combining multiple recommendation dimensions, and the accuracy of information recommendation for the target object is improved.
In some embodiments of the present invention, recommending the recommendation information set based on the recommendation parameter information, that is, the specific implementation process of S105 may include: S1051-S1052, as follows:
s1051, sorting the recommendation information set according to the recommendation parameter information to obtain a sorting result.
After the information recommending device obtains the recommendation parameter information, the information recommending device can judge the interest degree of the target object for each recommendation information in the recommendation information set according to the recommendation parameter information, and then determine the sequence for each recommendation information according to the interest degree of the target object for each recommendation information, so that the information recommending device can obtain the sequencing result.
And S1052, presenting the sequencing result to complete the recommendation of the recommendation information set.
And after the information recommendation device obtains the sequencing result, returning the sequencing result to the terminal of the target object, and displaying the sequencing result to the target object through a display interface of the terminal, so that the information recommendation device completes the information recommendation process.
In the embodiment of the invention, the information recommendation device can sequence the recommendation information sets to obtain a sequencing result, and the sequencing result is presented to the target object through the terminal, so that the information recommendation device can recommend the recommendation information sets based on the recommendation parameter information.
In some embodiments of the present invention, referring to fig. 6, obtaining recommendation characteristics corresponding to each recommendation information from each recommendation information in the recommendation information set and the historical browsing record, that is, a specific implementation process of S103, may include: S1031-S1033, as follows:
s1031, extracting at least one sparse feature and at least one dense feature according to each piece of recommendation information and each historical browsing record; sparse features are high-dimensional features requiring dimensionality reduction, dense features are directly obtained features.
When the information recommending device obtains the recommending characteristic corresponding to each piece of recommending information, aiming at the combination formed by each piece of recommending information and the historical browsing record, at least one high-dimensional characteristic, namely a sparse characteristic, and at least one characteristic which does not need dimension reduction, namely a dense characteristic are extracted and obtained from the combination.
In the embodiment of the present invention, the sparse features and the dense features may be extracted from the identifier, the category, and the tag of each piece of information browsed by the target object included in the history browsing record, and the operation and interaction of the target object on each piece of information, that is, the identifier information and the attribute information included in the recommendation information, by using a preset sparse model and a preset dense model.
In some embodiments of the present invention, the sparse features may include a feature and an information ID respectively corresponding to a category and a tag of information browsed by the target object, an information ID of recommendation information, and a feature respectively corresponding to an information category and an information tag, and may further include an ID of the target object, and other features having a high dimension and sparse values, which is not limited herein. The dense features may include word vectors corresponding to titles of information browsed by the target object, vectors corresponding to cover images of the browsed information, features corresponding to a table of recommendation information, and features corresponding to cover images of the recommendation information, and may also include other available dense features, which is not limited herein in the embodiments of the present invention.
It should be noted that, in the embodiment of the present invention, each sparse feature may be regarded as one of the feature sets corresponding to the sparse feature, and similarly, each dense feature may also be regarded as one of the feature sets corresponding to the dense feature.
For example, when a certain sparse feature obtained by the information recommendation device is a feature corresponding to the category "entertainment" of the information browsed by the history of the target object, the feature set corresponding to the sparse feature may be a set consisting of "entertainment", "current affairs", "military", "financial", and "life".
S1032, performing dimensionality reduction on the at least one sparse feature by using the at least one preset feature matrix to obtain at least one dimensionality reduction feature; the pre-set feature matrix consists of embedded vectors corresponding to the sparse features.
After obtaining the at least one sparse feature and the at least one dense feature, the information recommendation device needs to obtain a preset at least one preset feature matrix stored in advance, and obtain a dimension reduction feature corresponding to each sparse feature by using a preset feature matrix corresponding to a feature set to which each sparse feature in the at least one sparse feature belongs to reduce the dimension of the at least one sparse feature, so that the at least one dimension reduction feature can be obtained.
It should be noted that, in the embodiment of the present invention, the preset feature matrix is set in advance, and the preset feature matrix corresponds to the feature set to which the sparse feature belongs. Each row of the predetermined feature matrix may be regarded as an embedded vector, and the embedded vector corresponds to the sparse feature. Thus, the number of rows of each preset feature matrix is the same as the number of features in the feature set corresponding to the sparse feature. The number of columns of the preset feature matrix, namely the dimension of the embedded vector, is a super parameter which is set in advance.
And S1033, splicing the at least one dimension reduction feature and the at least one dense feature to obtain a recommended feature.
After the information recommendation device performs dimensionality reduction on the at least one sparse feature to obtain at least one dimensionality reduction feature, the at least one dimensionality reduction feature and the at least one dense feature can be spliced, and the result obtained by splicing is used as a recommendation feature so that the recommendation feature can be sent to a preset recommendation model for calculation subsequently.
It should be noted that the information recommendation device selects the splicing method according to the form of the dense features. When the dense features are vectors, the information recommendation device splices at least one dimension reduction feature and at least one sparse feature end to obtain a spliced vector, and the spliced vector is used as a recommendation feature; when the dense features are matrixes, the information recommendation device firstly judges whether the dimension of at least one dimension reduction feature is the same as the column number of at least one dense feature, if the dimension reduction feature is the same as the column number of the dense features, the dimension reduction feature and the dense features are spliced into a matrix according to the dimension reduction feature and the dense features and serve as recommendation features, and if the dimension reduction feature and the dense features are not the same, zero padding operation is utilized to splice the dimension reduction feature and the dense features into a matrix to obtain the recommendation features.
For example, referring to fig. 7, an embodiment of the present invention provides a schematic of obtaining recommended features, in fig. 7, an information recommendation device obtains N sparse features, which are respectively sparse feature 1, sparse feature 2, … …, and sparse feature N, and obtains 1 dense feature, that is, dense feature 1, then the information recommendation device performs a dimension reduction operation on sparse feature 1, sparse feature 2, … …, and sparse feature N, respectively obtains feature 1, dimension reduction feature 2, … …, and dimension reduction feature N, and then joins dimension reduction feature 1, dimension reduction feature 2, … …, dimension reduction feature N, and dense feature 1, so that the information recommendation device can obtain recommended features.
In the embodiment of the invention, the information recommendation device can extract at least one sparse feature needing dimension reduction and at least one dense feature needing no dimension reduction aiming at each recommendation information and each historical browsing record, then dimension reduction is carried out on the at least one sparse feature to obtain at least one dimension reduction feature, and the at least one dimension reduction feature and the at least one dense feature are spliced to obtain the recommendation feature, so that the information recommendation device can input the recommendation feature into a preset recommendation model for operation in the subsequent process.
In some embodiments of the present invention, for each piece of recommendation information and historical browsing record, extracting at least one sparse feature and at least one dense feature, that is, a specific implementation process of S1031 may include: s1031a-S1031c, as follows:
and S1031a, combining each piece of recommendation information with the historical browsing records respectively to obtain combination information corresponding to each piece of recommendation information.
When the information recommendation device obtains at least one sparse feature and at least one dense feature, each piece of recommendation information is combined with the historical browsing records to obtain combined information corresponding to each piece of recommendation information, so that the recommendation features can be obtained from the combined information in the follow-up process, and then recommendation scores are obtained, namely, each piece of recommendation information is scored by combining the historical browsing records of the target object.
And S1031b, extracting at least one sparse feature from the combined information corresponding to each piece of recommendation information.
After obtaining the combined information corresponding to each piece of recommendation information, the information recommendation device performs feature extraction on the combined information corresponding to each piece of recommendation information by using at least one pre-stored preset sparse model to obtain at least one sparse feature.
It can be understood that, in the embodiment of the present invention, there is a corresponding relationship between the preset sparse model and the sparse feature, for example, when extracting the sparse feature corresponding to the information category, the preset sparse model corresponding to the information category needs to be used for extraction, and when extracting the sparse feature corresponding to the information ID, the preset sparse model corresponding to the information ID needs to be used for extraction.
It should be noted that, in the embodiment of the present invention, the preset sparse model may be set according to the type of the sparse feature. For example, when digital content needs to be extracted as a sparse feature, for example, an information ID, a preset sparse model may be set as a regular expression to match the information ID from the historical browsing record and the recommendation information; when the content of the text type needs to be extracted as the sparse feature, such as the information category, a One-bit efficient encoder (One-Hot Encoding) can be used as the preset sparse model.
And S1031c, extracting at least one dense feature from the combined information corresponding to each piece of recommendation information.
Similarly, the information recommendation device extracts the combined information corresponding to each recommendation information by using at least one preset dense model stored in advance to obtain at least one dense feature.
It should be noted that the information recommendation apparatus may obtain the dense features by using a Word to Vector (Word 2Vec) model, and may also obtain the dense features by using a Bidirectional Encoder (BERT). Of course, other models may be used to obtain dense features, and embodiments of the present invention are not limited herein.
In the embodiment of the invention, the information recommendation device may combine each piece of recommendation information with the historical browsing records, and obtain at least one sparse feature and at least one dense feature respectively by using at least one preset sparse model and at least one dense model, so that the information recommendation device can process the at least one sparse feature and the at least one dense feature in the subsequent process to obtain the recommendation feature.
In some embodiments of the present invention, performing dimension reduction on at least one sparse feature by using at least one preset feature matrix to obtain at least one dimension reduction matrix, that is, a specific implementation process of S1032 may include: s1032a-S1032b, as follows:
s1032a, determining a preset feature matrix for each of the at least one sparse feature.
When the information recommendation device performs dimension reduction on at least one sparse feature to obtain at least one dimension reduction feature, determining a feature set to which the sparse feature belongs for each sparse feature in the at least one sparse feature, and then determining a preset feature matrix according to the feature set to which the sparse feature belongs, so that dimension reduction can be performed subsequently according to the preset feature matrix corresponding to the feature set of each sparse feature.
S1032b, determining an embedding vector corresponding to each sparse feature according to each sparse feature and the preset feature matrix, and obtaining at least one dimension reduction feature.
After obtaining the preset feature matrix corresponding to the feature set to which each sparse feature belongs, the information recommendation device may multiply each sparse feature by each preset feature matrix, and use the obtained multiplication result as the dimension reduction feature corresponding to each sparse feature.
It should be noted that, in the embodiment of the present invention, a process of multiplying the sparse feature by the preset feature matrix to obtain the dimension reduction feature may be regarded as a process of searching and extracting an embedded vector corresponding to the sparse feature from the preset feature matrix.
For example, when the sparse feature is a feature corresponding to an information category "entertainment", the information recommendation device may determine that the feature set to which the sparse feature belongs is a set consisting of "entertainment", "current affairs", "military", "financial", and "life", a preset feature matrix corresponding to the feature set to which the sparse feature belongs is a 5 × 4 matrix, a row 1 of the preset feature matrix is an embedded vector corresponding to the "entertainment" category, a row 2 of the preset feature matrix is an embedded vector corresponding to the "current affairs" category, a row 3 of the preset feature matrix is an embedded vector corresponding to the "military" category, a row 4 of the preset feature matrix is an embedded vector corresponding to the "financial" category, and a row 5 of the preset feature matrix is an embedded vector corresponding to the "life" category, at this time, a process of determining the dimension reduction feature may be regarded as a process of extracting the embedded vector of the row 1 of the preset feature matrix.
In the embodiment of the invention, the information recommendation device can determine the feature set to which each sparse feature belongs for each sparse feature, and then extract the embedded vector corresponding to each sparse feature from the preset feature matrix corresponding to the feature set to which each sparse feature belongs, so that the dimension reduction feature of each sparse feature is obtained, and the subsequent information recommendation device can obtain the recommendation feature according to the dimension reduction feature and the dense feature.
In some embodiments of the invention, the preset recommendation model comprises a shared layer and at least one independent layer; by using the preset recommendation model and the recommendation feature, calculating recommendation parameter information for each recommendation information, that is, a specific implementation process of S104 may include: S1041-S1042, as follows:
and S1041, performing vector conversion on the recommendation characteristics by using a sharing layer of a preset recommendation model to obtain an output vector of the sharing layer.
The information recommendation device takes the recommendation characteristics as input, and after the recommendation characteristics are sent into the preset recommendation model, the recommendation characteristics of each recommendation information are calculated by using a sharing layer of the preset recommendation model, namely, vector conversion is carried out on the recommendation characteristics, and the conversion result of the sharing layer is determined as the output vector of the sharing layer.
It should be noted that the preset sharing layer of the recommendation model refers to a fully connected layer formed by a plurality of hidden layers, and the parameters of the sharing layer are shared by the independent layer of the playing completion degree, the independent layer of the effective playing rate, the independent layer of the front feedback rate, and the independent layer of the screen interaction rate. The number of hidden layers in the shared layer and the number of neurons in each hidden layer may be set according to actual conditions, and the embodiment of the present invention is not specifically limited herein.
In some embodiments of the present invention, when the information recommendation apparatus calculates the recommendation feature by using the sharing layer in the preset recommendation model, the information recommendation apparatus substantially calculates the recommendation feature by using the hidden layer in the sharing layer. When the hidden layer is used for calculation, firstly, the input of the previous layer is multiplied by the weight parameter to obtain a product result, then the product result is added with the bias parameter of the cover layer to obtain a sum value result, then the sum value result is mapped between 0 and 1 by using a preset activation function to obtain the output of the hidden layer, and the process is circulated until the last hidden layer is calculated, and an output vector is obtained by the output of the last hidden layer.
For example, the embodiment of the present invention provides a calculation formula of a hidden layer, as shown in formula (1):
a l＝σ((w l) Ta l-1+b l) (1)
wherein, w lIs a weight parameter of the l-th layer, b lIs a weight bias parameter of the l-th layer, a lIs the output of the l-th layer, a l-1σ is the activation function for the output of layer l-1.
Information recommendation device obtaining output a of l-1 layer l-1Then, the weighting parameter w of the l-th layer is combined lWeight bias parameter b of the l-th layer lOutput of the l-th layer a lActivation functions σ and a l-1In the formula (1), the output of the l-th layer is calculated.
Illustratively, the embodiment of the present invention gives an illustration of a shared layer, as shown in fig. 8, the shared layer has three hidden layers, i.e., L1, L2, and L3, L1 has 3 neurons, L2 has 4 neurons, and L3 has 2 neurons. The recommended features are respectively input into 3 neurons of L1, after 3 neurons of L1 respectively calculate the recommended features, the calculation results are used as input and are respectively sent into 4 neurons of L2, 4 neurons of L2 respectively continue to calculate the calculation results of the 3 neurons of L1, then the calculation results are respectively input into 2 neurons of L3 for calculation, 2 neurons of L3 can obtain 2 vectors after the calculation is completed, at this time, the 2 vectors can be spliced together by the sharing layer to be used as output vectors, and the 2 vectors can also be used as output vectors to be sent into the independent layer for calculation in the subsequent steps.
S1042, respectively calculating a recommended parameter of an output vector by using each independent layer of at least one independent layer to obtain recommended parameter information; wherein the parameters of each individual layer are independent of each other.
After obtaining the output vector, the information recommendation transposing calculates the output vector by using at least one of an independent layer of playing completion degree, an independent layer of effective playing rate, an independent layer of front feedback rate and an independent layer of screen interaction rate in a preset recommendation model, so that recommendation parameter information corresponding to each recommendation information, namely at least one of playing completion degree, effective playing rate, front feedback rate and screen interaction rate, can be obtained.
It should be noted that the independent layer of the playing completion degree, the independent layer of the effective playing rate, the independent layer of the front feedback rate, and the independent layer of the screen interaction rate may be all connected layers composed of a plurality of hidden layers, but these all connected layers are independent from each other and do not share parameters.
It can be understood that the number of hidden layers in each independent layer and the number of neurons in each hidden layer may be the same or different, and the number of hidden layers in each independent layer and the number of neurons in each hidden layer may be set according to actual situations, which is not limited herein in the embodiment of the present invention.
It should be noted that the calculation method of the hidden layer in each independent layer is the same as the calculation method of the hidden layer in the shared layer in S1041, and details are not repeated here.
Illustratively, an embodiment of the present invention provides a schematic of a preset recommendation model, as shown in fig. 9, an output of a shared layer is respectively connected to an independent layer PCR-FC of a playing completion degree, an independent layer CTR-FC of an effective playing rate, an independent layer PFR-FC of a front feedback rate, and an independent layer SIR-FC of a screen interaction rate. The information recommendation device calculates recommendation characteristics through the sharing layer, after the output vectors are obtained, the output vectors are respectively input into the PCR-FC, the CTR-FC, the PFR-FC and the SIR-FC, the independent layers respectively calculate the output vectors to obtain the playing completion degree, the effective playing rate, the front feedback rate and the screen interaction rate, and accordingly the recommendation information can be scored according to the obtained parameters according to the information recommendation parameters in the follow-up process.
In the embodiment of the present invention, the information recommendation device may calculate the recommendation characteristic through a sharing layer of a preset recommendation model, and then calculate the output of the sharing layer by using each of at least two independent layers to obtain at least one of a playing completion degree, an effective playing rate, a front feedback rate, and a screen interaction rate, so that a subsequent information recommendation device scores recommendation information according to the obtained parameters.
In some embodiments of the present invention, according to the recommendation parameter information, the recommendation information sets are sorted to obtain a sorting result, that is, a specific implementation process of S1051 may be S1051a-S1051b, as follows:
s1051a, calculating a recommendation score for each recommendation information according to the recommendation parameter information; the recommendation score characterizes the interest degree of the target object in each recommendation information.
After the information recommendation device obtains the recommendation parameter information, the interest degree of a target object in each recommendation information can be determined according to at least one of the playing completion degree, the effective playing rate, the positive feedback rate and the screen interaction rate in the recommendation parameter information, namely, the interest degree of the target object in each recommendation information is scored, and a recommendation score is obtained.
It should be noted that, because the playing completion degree, the effective playing rate, the positive feedback rate, and the screen interaction rate may all represent the interest degree of the target object in the recommendation information, the information recommendation device may calculate the recommendation score by using a part of parameters of the playing completion degree, the effective playing rate, the positive feedback rate, and the screen interaction rate, or calculate the recommendation score by using all parameters of the playing completion degree, the effective playing rate, the positive feedback rate, and the screen interaction rate, which is not limited in this embodiment of the present invention.
For example, the information recommendation device may use only any one of the play completion degree, the effective play rate, the positive feedback rate, and the screen interaction rate as the recommendation score, may calculate the recommendation score by using the play completion degree and the effective play rate, and may calculate the recommendation score by using the play completion degree, the effective play rate, the positive feedback rate, and the screen interaction rate in a comprehensive manner.
It should be noted that, when the information recommendation device calculates the recommendation score by using a part of parameters or all of the parameters in the playing completion degree, the effective playing rate, the positive feedback rate and the screen interaction rate, the parameters may be accumulated to obtain the recommendation score of each piece of recommendation information, the parameters may also be accumulated to obtain the recommendation score of each piece of recommendation information, and other operations may also be performed on the parameters to obtain the recommendation score of each piece of recommendation information.
It should be noted that, since the value ranges of the play completion, the effective play rate, the positive feedback rate, and the screen interaction rate are all between [0 and 1], when at least one recommendation parameter is used for accumulation, the accumulation result may exceed 1, and therefore, the information recommendation device may normalize the accumulation result, and take the final normalized result as the recommendation score.
S1051b, sorting the recommendation information set by using the recommendation score to obtain a sorting result.
After the information recommending device calculates the recommending scores, the information recommending device can determine the sequence for each piece of recommended information according to the recommending scores from high to low, so that the information recommending device obtains the sequencing result aiming at the recommended information set, the sequencing result is conveniently returned to the terminal in the follow-up process, and the sequencing result is presented to the target object through the terminal.
In the embodiment of the invention, the information recommendation device can score each piece of recommendation information according to at least one of the playing completion degree, the effective playing rate, the positive feedback rate and the screen interaction rate in the recommendation parameter information to obtain the recommendation score, and then sort the recommendation information set according to the recommendation score to obtain the sorting result, so that the information recommendation device can evaluate the recommendation information by using multiple dimensions, and the accuracy of information recommendation is improved.
In some embodiments of the present invention, obtaining a recommendation information set for a target object according to a historical browsing record of the target object, that is, a specific implementation process of S102 may include: S1021-S1023, as follows:
and S1021, acquiring a category label corresponding to each browsing record in the historical browsing records.
The information recommendation device obtains the category label corresponding to the information browsed by the target object from the historical browsing record of the target object, namely the category label of each browsing record, so as to conveniently count the commonly used category labels subsequently.
And S1022, counting the common category labels of the target object from the category labels corresponding to each browsing record.
After the information recommendation device obtains the category label corresponding to each browsing record, the information recommendation device can count the category label corresponding to each browsing record to obtain the category label with the largest browsing frequency of the target object, and use the category label as a common category label.
For example, an illustration of counting common category labels is provided in the embodiment of the present invention, as shown in fig. 10, a target object has 5 browsing records, which are browsing record 1, browsing record 2, browsing record 3, browsing record 4, and browsing record 5. Wherein, the time of occurrence of the browsing record 1 is 21: 30, the category label corresponding to the event is "art", the time of occurrence of the browsing record 2 is 20:00, the category label corresponding to the event is "series", the time of occurrence of the browsing record 3 is 16:50, the category label corresponding to the event is "art", the time of occurrence of the browsing record 4 is 12:03, the category label corresponding to the event is "cartoon", the time of occurrence of the browsing record 5 is 09:10, and the category label corresponding to the event is "fresh", at this time, the information recommendation device may count the category label "art" with the largest number of times of browsing the target object from the 5 browsing records, and use the category label as a commonly used category label.
And S1023, screening the full information set according to the common class labels, and determining the recommendation information set of the target object.
After obtaining the common category labels, the information recommendation device compares the category label of each piece of information in the total information set with the common category label, and selects out the information with the same category label as the common category label in the total information set to obtain a recommendation information set.
It is understood that the full information set in the embodiment of the present invention refers to a set composed of all information stored in the database.
In the embodiment of the invention, the information recommendation device can determine the common category label of the target object from the historical browsing record of the target object, and screen out the recommendation information set from the total information set by using the common category label, so that the information recommendation device can directly score the recommendation information in the recommendation information set subsequently, and the operation amount of the information recommendation device is reduced.
In some embodiments of the present invention, in addition to the effective play rate, the play completion, the positive feedback rate and the screen interaction rate, the recommendation parameter information may further include at least one of a play duration and comment information; the playing duration represents the time length of the target object for playing the recommendation information, and the comment information represents the probability of the target object commenting on the recommendation information.
It can be understood that the playing duration is the predicted time length that the target object may play the recommendation information, and the longer the target object plays the recommendation information, the more interested the target object is in the recommendation information; the comment information is the predicted probability that the target object comments on the recommendation information, and the higher the probability that the target object comments on the recommendation information, the higher the attention degree of the target object to the recommendation information is, that is, the more interesting the target object is to the recommendation information.
In the embodiment of the invention, the recommendation parameter information can also comprise at least one of playing time and comment information, so that the information recommendation device can perform information recommendation by combining the playing time, user comments and other recommendation dimensions capable of reflecting the interest of the target object besides performing information recommendation by using at least one of the effective playing rate, the playing completion degree, the positive feedback rate and the screen interaction rate, and further improve the accuracy of information recommendation for the target object.
In some embodiments of the present invention, before receiving an information browsing request of a target object and acquiring a history browsing record of the target object according to the information browsing request, that is, before S101, the method may further include: S106-S110, as follows:
s106, at least one temporary model is constructed by utilizing at least one group of hyper-parameters and the initial recommendation model.
Before receiving an information browsing request of a target object, an information recommendation device needs to obtain a trained preset recommendation model. The information recommendation device needs to construct a temporary model corresponding to each group of hyper-parameters by using at least one group of hyper-parameters and an initial recommendation model to obtain at least one temporary model, and in the subsequent step, the temporary models can be trained so as to select a model with the best effect as a preset recommendation model.
The hyper-parameters are parameters that need to be set before the model training is started. The hyper-parameters may include a learning rate and a step length, and may further include the number of hidden layers in the fully connected layer and the number of neurons in the hidden layers. The specific hyper-parameter may be set according to actual requirements, and the embodiment of the present invention is not limited herein.
S107, training at least one temporary model by using training sample data to obtain at least one loss function value corresponding to at least one independent layer; the training sample data comprises a training browsing record and a training recommendation information set.
The information recommendation device obtains training sample data from a storage area of the information recommendation device, then the training sample data are respectively used as input and are respectively sent into at least one temporary model to be trained, and a corresponding loss function value in at least one independent layer in each temporary model is obtained, so that the information recommendation device can obtain at least one loss function value.
It can be understood that the training sample includes browsing records for training, i.e., training browsing records, and a recommendation information set for training, i.e., a training recommendation information set.
It should be noted that the training sample may be composed of historical browsing records of all users and sets of recommendation information for all users stored in the database, or may be composed of historical browsing records of some users and sets of recommendation information for some users stored in the database.
And S108, obtaining current training parameters respectively corresponding to at least one temporary model.
The information recommendation device obtains each temporary model in the at least one temporary model, and the weight of each independent layer in the at least one independent layer is used as the current training parameter when the current training round is carried out, so that the current training parameter can be updated conveniently.
S109, combining the at least one loss function value and the current training parameters respectively corresponding to the at least one temporary model, and adjusting the current training parameters respectively corresponding to the at least one temporary model according to a combination result to obtain at least one group of training parameters.
The information recommendation device combines the at least one loss function value and the current training parameters corresponding to each of the at least one temporary model to obtain a combined result, adjusts and updates the current training parameters of each of the at least one temporary model according to the combined result, and takes the adjusted current training parameters as at least one group of training parameters.
It should be noted that, combining the at least one loss function value with the current training parameter corresponding to each of the at least one temporary model means combining the weight of each of the at least one independent layer of each of the temporary models with the at least one loss function value to obtain an overall loss function value of each of the temporary models, and using the overall loss function value of each of the temporary models as a combination result.
S110, forming at least one intermediate recommendation model by using at least one group of training parameters and at least one temporary model, and selecting a preset recommendation model from the at least one intermediate recommendation model by using test sample data.
After the information recommendation device obtains at least one group of training parameters, each group of training parameters in the at least one group of training parameters and each temporary model in the at least one temporary model are correspondingly used for forming an intermediate recommendation model to obtain at least one intermediate recommendation model, then each intermediate recommendation model in the at least one intermediate recommendation model is used for calculating test sample data to obtain at least one test accuracy, and finally the intermediate recommendation model with the highest test accuracy is selected to serve as a final preset recommendation model, so that the information recommendation device completes training of the preset recommendation model.
In the embodiment of the invention, the information recommendation device can set a plurality of groups of hyper-parameters to obtain a plurality of temporary models, then training a plurality of temporary models respectively by using the training sample data to obtain a loss function value corresponding to at least one independent layer in each temporary model, and a weighted combination is performed using the at least one loss function value and the current training parameters for each of the provisional models, updating the current training parameters according to the combination result to obtain at least one group of training parameters and further obtain at least one intermediate recommendation model, and using the test sample data to select an intermediate temporary model with the best effect from at least one intermediate recommended model obtained by training as a preset training model, thus, through the mode of grid search, the information recommendation device can obtain a preset training model with a good effect, and therefore the accuracy of information recommendation is improved.
In some embodiments of the present invention, training at least one temporary model by using training sample data to obtain at least one loss function value corresponding to at least one independent layer, that is, a specific implementation process of S107 may include: S1071-S1073, as follows:
s1071, aiming at training sample data, determining training recommendation characteristics; training sample data is provided with a training label; the training labels comprise at least one of a playing completion label, an effective playing rate label, a front feedback rate label and a screen interaction rate label.
The information recommending device determines the recommended features corresponding to the training samples from the training sample data, namely the training recommended features, so that the follow-up information recommending device can predict the training recommended features by using the temporary model.
It can be understood that, since the information recommendation device needs to predict at least one of the playing completion degree, the effective playing rate, the positive feedback rate, and the screen interaction rate of the training sample data through the temporary model, at least one of the playing completion degree label, the effective playing rate label, the positive feedback rate label, and the screen interaction rate label may be included in the training sample data, so that the information recommendation device performs supervised training according to the training sample data with the label.
It should be noted that, a process of determining the training recommended features from the training sample data is similar to the process of S1031 to S1033, and details of the embodiment of the present invention are not described here again.
S1072, predicting the training recommendation characteristics by using at least one temporary model to obtain a prediction result; the prediction result comprises at least one of a prediction playing completion degree, a prediction effective playing rate, a prediction positive feedback rate and a prediction screen interaction rate.
The information recommendation device predicts the training recommendation characteristics by using each temporary model in at least one temporary model to obtain at least one group of prediction results, wherein each group of prediction results comprises at least one of a predicted playing completion degree, a predicted effective playing rate, a predicted positive feedback rate and a predicted screen interaction rate.
It should be noted that a specific process of predicting the training recommended features to obtain a prediction result is similar to S1041-S1042, and details are not repeated herein in the embodiments of the present invention.
S1073, obtaining at least one loss function value corresponding to at least one independent layer by using the prediction result and the label of the training sample data.
The information recommendation device can calculate the error of the prediction result and the training label according to the prediction result, namely at least one of the predicted playing completion degree, the predicted effective playing rate, the predicted positive feedback rate and the predicted screen interaction rate, and the label of the training sample data, namely at least one of the playing completion degree label, the effective playing rate label, the positive feedback rate label and the screen interaction rate label, so that at least two loss function values corresponding to at least one independent layer can be obtained.
It should be noted that, since different independent layers are trained for different learning objectives, an appropriate loss function needs to be selected for each independent layer according to the characteristics of the learning objectives. The independent layer of the playing completion degree needs to predict the information playing progress, belongs to a regression task, and can select a loss function suitable for the regression task as the loss function of the independent layer of the playing completion degree; the independent layer of the effective playing degree predicts the probability of the information being effectively played, belongs to a two-classification task, and can be suitable for the loss function of the two-classification task as the loss function of the independent layer of the effective playing degree; similarly, the independent layer of the positive feedback rate and the independent layer of the screen interaction rate respectively predict the probability that the information has positive feedback and the probability that the information has screen interaction in playing, belong to a binary task, and can select a loss function suitable for the binary task. As a loss function of the independent layer of the front feedback rate and the independent layer of the screen interaction rate, respectively.
In some embodiments of the present invention, a square loss function may be selected for the independent layer of the playing completion degree, that is, a difference is first made between the tag of the playing completion degree and the predicted playing completion degree, and then a square is obtained for a difference result to obtain a loss function value of the independent layer of the playing completion degree.
For example, the embodiment of the present invention provides a formula for calculating the prediction error of the independent layer of the playing completion, as shown in formula (2),
loss PCR＝(y PCR-PPCR) 2(2)
wherein, y PCRFor the tag of the playing completion, PPCR is the predicted playing completion, loss PCRIs the loss value of the independent layer of the playing completion. After calculating the predicted playback completion PPCR, the information recommendation device will compare the PPCR with the tag y of the playback completion PCRSubstituting the formula (2) for calculation to obtain the loss function value loss of the independent layer of the playing completion degree PCR。
In some embodiments of the present invention, a cross entropy loss function may be selected for the independent layer of the effective playing rate, that is, 1 is used to make a difference with the tag of the effective playing rate and the predicted effective playing rate, respectively, a logarithm is obtained for a difference between 1 and the predicted effective playing rate to obtain a logarithm result, the difference between 1 and the tag of the effective playing rate is multiplied by the logarithm result to obtain a first product result, then the logarithm of the predicted effective playing rate is multiplied by an opposite number of the tag of the effective playing rate to obtain a second product result, and finally the second product result is subtracted from the first product result to obtain a loss function value of the independent layer of the effective playing rate.
For example, the embodiment of the present invention provides a formula for calculating the prediction error of the independent layer of the effective play-back rate, as shown in formula (3):
loss CTR＝-y CTR×log(PCTR)-(1-y CTR)×log(1-PCTR) (3)
wherein, y CTRIs the effective Play Rate Label, PCTR is the predicted effective Play Rate, loss CTRThe prediction error of the independent layer for the effective play-out rate. After obtaining the predicted effective play rate PCTR, the information recommendation device labels y the effective play rate CTRAnd PCTR into equation (3) to calculate the loss function value loss of the independent layer for the effective play rate CTR。
In some embodiments of the present invention, an optimized cross entropy loss function may be selected for the independent layer of the front feedback rate, that is, when the label of the front feedback rate is 1, first, a difference is made between 1 and the predicted front feedback rate, a classification sample weight power is solved for a difference result to obtain a first-time result, and then the first-time result, an inverse number of the label category weight, and a logarithm of the predicted front feedback rate are multiplied to obtain a prediction error of the independent layer of the front feedback rate when the label of the front feedback rate is 1; when the label of the front feedback rate is 0, solving the classification sample weight power of the predicted front feedback rate to obtain a second power result, and multiplying the inverse number of the difference value between 1 and the label class weight, the second power result and the logarithm of the difference value between 1 and the predicted front feedback rate to obtain the loss function value of the independent layer of the front feedback rate when the label of the front feedback rate is 0.
Illustratively, embodiments of the present invention provide a loss function of the independent layer of the positive feedback rate, as shown in equation (4):
Figure BDA0002242561060000311
wherein α is the label class weight, γ is the classified sample weight, y PFRThe information recommending means obtains label category weight α, classification sample weight gamma, label y of positive feedback rate PFRAfter the predicted front feedback rate PPFR, the above parameters may be substituted into the formula (4) for calculation to obtain the loss function value loss of the independent layer of the front feedback rate PFR。
In some embodiments of the invention, the cross-entropy loss may be optimized for selection of an independent layer of screen interactivity. When the label of the screen interaction rate is 1, firstly, subtracting the screen interaction rate from the predicted screen interaction rate by 1, solving the classification sample weight power of the difference value between the screen interaction rate and the predicted screen interaction rate to obtain a third power result, and then multiplying the third power result, the opposite number of label class weights and the logarithm of the predicted screen interaction rate by each other to obtain the prediction error of the independent layer of the screen interaction rate when the label of the screen interaction rate is 1; when the label of the screen interaction rate is 0, solving the classification sample weight power of the prediction screen interaction rate to obtain a fourth power result, and multiplying the fourth power result by the inverse number of the difference value between 1 and the label class weight and the logarithm of the difference value between 1 and the prediction screen interaction rate to obtain the loss function value of the independent layer of the screen interaction rate when the label of the screen interaction rate is 0.
Illustratively, the embodiment of the present invention provides a formula for calculating a prediction error of an independent layer of a screen interaction rate, where the predicted screen interaction rate is calculated by multiplying a temporary screen interaction rate and a predicted effective play rate, as shown in formula (5):
Figure BDA0002242561060000321
wherein α is the label class weight, γ is the classified sample weight, y SIRIs a screen interaction rate label, PSIR is a temporary screen interaction rate, PCTR is a predicted effective play rate, PSIR multiplied by PCTR is a predicted screen interaction rate, loss SIRThe information recommending device obtains label class weight α, classification sample weight gamma and screen interaction rate label y SIRAfter predicting the effective playback rate PCTR, the temporary screen interaction rate PSIR may substitute the above parameters in equation (5), and may obtain the loss function value loss of the independent layer of the screen interaction rate SIR。
In the embodiment of the invention, the information recommendation device can train at least one temporary model by using the training sample data with the label to obtain at least one loss function value corresponding to at least one independent layer, so that the current training parameter can be adjusted according to the at least one loss function value subsequently.
In some embodiments of the present invention, when the prediction result includes the predicted playing completion, the predicted effective playing rate, the predicted positive feedback rate, and the predicted screen interaction rate, the training recommendation feature is predicted by using at least one temporary model to obtain the prediction result, that is, the specific implementation process of S1072 may include: s1072a-S1072d, as follows:
s1072a, predicting the training recommendation characteristics by using at least one temporary model to obtain a temporary playing completion degree, a predicted effective playing rate, a predicted positive feedback rate and a temporary screen interaction rate.
When the information recommendation device predicts the training recommendation characteristics by using at least one temporary model, the temporary playing completion degree, the predicted effective playing rate, the predicted positive feedback rate and the temporary screen interaction rate can be obtained by outputting, so that the information recommendation device can calculate the predicted playing completion degree and the predicted screen interaction rate according to the four parameters.
It should be noted that, because the playing completion and the screen interaction rate are conditional probabilities under the playing behavior, there inevitably exist samples that are not played in the training sample data, the labels of the playing completion and the screen interaction rate are not uniformly distributed, and the screen interaction behavior is relatively sparse in the playing samples. In order to avoid the influence of the unplayed samples on model training, the information recommendation model can calculate the predicted playing completion and the predicted screen interaction rate according to the temporary playing completion, the predicted effective playing rate, the predicted front feedback rate and the temporary screen interaction rate, so that the independent layer of the playing completion and the independent layer of the screen interaction rate can be concentrated on the task target of the information recommendation model when the loss function is calculated.
S1072b, determining the predicted playing completion degree according to the temporary playing completion degree and the predicted effective playing rate.
The information recommending device can multiply the temporary playing completion degree and the predicted effective playing rate, and the obtained product is used as the predicted playing completion degree, so that the information recommending device can associate the playing completion degree with the effective playing rate, and the influence of an unplayed sample on the playing completion degree is reduced.
S1072c, determining the predicted screen interaction rate according to the predicted effective play rate and the temporary screen interaction rate.
Similarly, the information recommending device multiplies the predicted effective playing rate by the temporary screen interaction rate, and the obtained product result is used as the predicted screen interaction rate, so that the information recommending device associates the screen interaction rate with the effective playing rate and reduces the influence of the unplayed sample on the screen interaction rate.
S1072d, obtaining a prediction result by using the predicted playing completion degree, the predicted effective playing rate, the predicted front feedback rate and the predicted screen interaction rate.
After the information recommendation device calculates the predicted playing completion degree and the predicted screen interaction rate, the predicted front feedback rate and the predicted effective playing rate, and the calculated predicted playing completion degree and the predicted screen interaction rate are used together to form a prediction result, so that parameter updating can be performed according to the prediction result in the following process.
In the embodiment of the invention, the information recommendation model can respectively associate the effective play rate with the play completion rate and the screen interaction rate in a multiplication mode in the training process, so that the influence of the unplayed sample on the play completion rate and the screen interaction rate is reduced, the training effect of the temporary model is improved, the performance of the preset recommendation model is improved, and the accuracy of information recommendation is higher.
In some embodiments of the present invention, referring to fig. 11, combining the at least one loss function value and the current training parameters respectively corresponding to the at least one temporary model, and adjusting the current training parameters respectively corresponding to the at least one temporary model according to the combination result to obtain at least one set of training parameters, that is, a specific display process of S109 may include: S1091-S1094, as follows:
s1091, conducting weighted combination on the current training parameters corresponding to the at least one loss function value and the at least one temporary model respectively to obtain a combination result.
The information recommendation device weights the at least one loss function value and the weight of each independent layer in the current training parameter corresponding to the at least one temporary model respectively, and takes the overall loss function value obtained by weighting as a combination result.
In some embodiments of the present invention, when at least one of the loss function values includes a loss function value of an independent layer of the playing completion degree, a loss function value of an independent layer of the effective playing rate, a loss function value of an independent layer of the front feedback rate, and a loss function value of an independent layer of the screen interaction rate, a product of the loss function value of the independent layer of the playing completion degree and a weight corresponding to the independent layer of the playing completion degree, a product of the loss function value of the independent layer of the effective playing rate and a weight corresponding to the independent layer of the effective playing rate, a product of the loss function value of the independent layer of the front feedback rate and a weight corresponding to the independent layer of the front feedback rate, and a product of the loss function value of the independent layer of the screen interaction rate and a weight corresponding to the independent layer of the screen interaction rate may be accumulated to obtain an accumulation result.
For example, an embodiment of the present invention provides a formula for calculating an overall loss function value of a temporary model, as shown in formula (6):
loss＝a 1loss PCR+a 2loss CTR+a 3loss SIR+a 4loss PFR(6)
wherein, a 1Weight, loss, corresponding to an independent layer of playing completion PCRLoss function value of independent layer for playing completion, a 2Weight, loss, corresponding to an independent layer of effective play rate CTRLoss function value of independent layer of effective play rate, a 3Weight, loss, corresponding to an independent layer of screen interaction rate SIRLoss function value of an independent layer of screen interaction rate, a 4Weight, loss, corresponding to an independent layer for a positive feedback rate PFRThe loss function value of the individual layer for the positive feedback rate. After obtaining the parameters, the information recommendation device may substitute the parameters into equation (6) to obtain the overall loss function value loss.
S1092, respectively updating the weight corresponding to each independent layer in the at least one independent layer according to the combination result to obtain the updated weight corresponding to each independent layer.
Since the combined information corresponding to each temporary model is the weight of the loss function value of each independent layer, in order to train each temporary model to learn sufficiently during training, the information recommendation device needs to update the weight of each independent layer by using the combined information to obtain the updated weight corresponding to each independent layer, so that the information recommendation device can adaptively adjust the weight of the prediction error of each independent layer during the training process, and each independent layer can achieve a better training effect.
In some embodiments of the present invention, the corresponding weight of each independent layer may be updated according to the difference between the combination result and the training label.
When updating the weight, the information recommendation device may multiply the learning rate by the difference between the combination result and the training label, and use the product as the updated weight. Of course, the information recommendation device may also calculate the updated weight in other manners, and the embodiment of the present invention is not limited herein.
For example, an embodiment of the present invention provides a formula for calculating updated weights, as shown in formula (7):
w＝η×δ (7)
after obtaining the difference value δ between the combination result and the training label, the information recommendation device can substitute the learning rate η and the difference value δ between the combination result and the training label into equation (7) to calculate the updated weight w.
It can be understood that the information recommendation device not only updates the weight of each independent layer, but also reversely propagates the loss function value of each independent layer after obtaining the loss function value corresponding to each independent layer, and updates the weight parameter and the bias parameter of the hidden layer in each independent layer, that is, the parameter of each independent layer to obtain the updated parameter of each independent layer. Similarly, the information recommendation device updates the parameters in the sharing layer according to the utilization combination result to obtain the updated parameters of the sharing layer.
S1093, forming the adjusted current training parameters corresponding to each temporary model in the at least one temporary model by using the updated weight of each independent layer.
The information recommendation device uses the updated weight of each independent layer in the at least one independent layer of each temporary model to form the adjusted current training parameter corresponding to each temporary model, so that the information recommendation device can complete the adjustment process of the current training parameter of each temporary model in the at least one temporary model.
S1094, combining the adjusted current training parameters corresponding to each temporary model to obtain at least one group of training parameters.
After obtaining the adjusted current training parameters corresponding to each temporary model, the information recommendation device uses the adjusted current training parameters as the training parameters corresponding to each temporary model, so that the information recommendation device can obtain at least one set of training parameters corresponding to at least one temporary model, and then at least one intermediate recommendation model is formed by using at least one set of training parameters and at least one temporary model.
It should be noted that, in some embodiments of the present invention, in addition to obtaining at least one set of training parameters by using the adjusted current training parameters, the information recommendation apparatus may also use the adjusted current training parameters, the updated parameters of the shared layer, and the updated parameters of each independent layer to jointly form at least one set of training parameters, which is not limited herein.
In the embodiment of the invention, the information recommendation device can use at least one loss function value and the current training parameter for weighting to obtain an overall loss function value of at least one temporary model, the overall loss function value is used as a combination result, the weight of each independent layer is subjected to parameter updating according to the combination result to obtain the updated weight of each independent layer, the adjusted current training parameter corresponding to each temporary model is obtained according to the updated weight of each independent layer, and at least one group of training parameters are further obtained.
In the following, an exemplary application of the embodiment of the present invention in an actual application scenario will be described by taking a short video recommendation as an example.
Before recommending short videos, a recommendation system needs to train to obtain a video recommendation model. The video recommendation model mainly models four targets of playing completion, effective playing rate, positive feedback rate and screen interaction rate. The playing completion degree is a regression task, and the effective playing rate, the front feedback rate and the screen interaction rate are all classification tasks. The training samples of the video recommendation model are full exposure data before short video recommendation, namely, short videos historically browsed by all users, and display data provided by the recommendation system for all users, namely, short videos recommended to all users. Each sample in the training samples takes < playing completion degree, whether playing is effective, whether screen interaction operation exists, and whether forward feedback operation exists > as a label, wherein the playing completion degree is a continuous value and is equal to the proportion of the playing time to the total short video time, the value range is [0, 1], the effective playing rate, the forward feedback rate and the screen interaction rate are binary values, and the value range is 0 or 1. Whether effective playing is carried out or not can be determined by the playing time length and the playing completion degree, for short videos below 15s, effective playing can be defined as the playing time length being greater than 7s, and for short videos above 15s, effective playing can be defined as the playing completion degree being greater than 50%. Whether screen interaction exists or not refers to whether operations such as full screen, horizontal screen, fast forward, backward and the like exist in the short video playing process, and the operations can represent that a user has stronger watching interest in the short video. Whether the user has positive feedback refers to whether the user conducts operations such as approval, comment, collection or forwarding and the like on the short video, and the operations represent the spreading and interaction willingness of the user on the short video.
Referring to fig. 12, the model is a video recommendation model obtained by modeling the recommendation system according to four objectives of the playing completion degree, the effective playing rate, the positive feedback rate and the screen interaction rate, and the model further includes an input layer 1, a sharing layer 2, an independent layer 3 and an error calculation layer 4, wherein for convenience of understanding, the processes of extracting features and splicing the recommendation system are displayed in the input layer 1 of the video recommendation model. In the input layer 1, the recommendation system extracts n +1 features from the full exposure data, wherein the feature 1, the feature 2, … …, the feature n are all sparse features, and the feature n +1 is dense features. The sparse features comprise user ID, user video category interest, user video label interest, user historical playing video ID, video category and video label, and have the characteristics of high dimension and sparse value, and dimension reduction is needed to obtain sufficient representation information. The dense features refer to features directly extracted by a preset feature extraction model, and comprise video title vectors and video cover picture vectors. The recommendation system performs dimension reduction on the sparse features by using a preset dimension reduction matrix to respectively obtain an embedded vector 1 of the feature 1, an embedded vector 2 of the feature 2, … … and an embedded vector n of the feature n, and then splices the embedded vector 1, the embedded vectors 2, … …, the embedded vector n and the feature n +1 to obtain an input vector of the sharing layer 2. The full connection layer in the sharing layer 2 is composed of a plurality of hidden layers, the number of the hidden layers is a model hyper-parameter, and an optimal solution can be found by utilizing grid search. And each hidden layer is calculated according to the formula (1) to obtain an output vector. All parameters of the shared layer are shared among the four targets, i.e. loss of each target affects the update of the layer parameters. The independent layer 3 receives the output vector of the shared layer 2, and the fully connected layer of each target calculates the output vector according to the formula (1) to obtain the final predicted playing completion PPCR, the predicted effective playing rate PCTR, the predicted screen interaction rate PSIR and the predicted front feedback rate PPFR. In the independent layer 3, the fully-connected layer and the parameters of each target are independent of each other, and only the learning of the target is focused on, and the influence of other targets is avoided.
The error calculation layer 4 calculates the loss function and gradient of each target mainly according to the above prediction results, thereby updating the parameters of the independent layer 3 and the shared layer 2. As shown in equations (8) and (9), since the playing completion and the screen interaction rate are conditional probabilities in the playing behavior, when the full exposure data is used as the training sample, learning two joint probabilities of P (playing completion, | x) and P (screen interaction, | x) actually results in complication of parameter learning in the training process.
P (play-completed, | x) ═ P (play-completed | play) × P (play | x) (8)
P (screen interactive, play | x) ═ P (screen interactive | play) × P (play | x) (9)
In order to make the playing completion degree and the screen interaction rate focus on the learning of the target P (playing completion, playing | x) and P (screen interaction, playing | x), the recommendation system takes PCTR PPCR as the final predicted value of the playing completion degree and PCTR PSIR as the final predicted value of the screen interaction rate when calculating the loss function, so as to reduce the influence of the unplayed samples on the playing completion degree and the screen interaction rate. The recommendation system selects square loss as a loss function for the playing completeness, selects cross entropy loss as a loss function for the effective playing rate, and selects optimized cross entropy loss as a loss function for both the front feedback rate and the screen interaction rate, so that the loss functions are utilized to obtain an overall loss function by weighting. The recommendation system may then begin training the video recommendation model.
In the training process, the recommendation system not only updates the parameters of the independent layer and the parameters of the shared layer, but also updates the weight ratios a1, a2, a3 and a4 of the loss function of the playing completion degree, the loss function of the effective playing rate, the loss function of the positive feedback rate and the loss function of the screen interaction rate in the independent layer, so that the recommendation system can adaptively adjust the weight ratios of the loss functions between different targets.
After the training of the video recommendation model is completed, the recommendation system can use the video recommendation model to recommend short videos for the user. Referring to fig. 13, after the recommendation system receives a short video browsing request 1 from a user a, the recommendation system determines a recommended short video list according to categories and tags of short videos frequently browsed by the user a in the past, where the recommended short video list includes 2 videos, namely a short video a and a short video b. And then, the recommendation system combines the historical browsing records 3 of the user A obtained from the database with the short videos a and b respectively to obtain recommendation characteristics a corresponding to the short videos a and recommendation characteristics b corresponding to the short videos b, inputs the recommendation characteristics a and the recommendation characteristics b into the video recommendation model 4 for scoring, sorts the short videos a and the short videos b according to the scores, and returns the sorting result to the user A. Meanwhile, the recommendation system can be updated by the sequencing result, so that the recommendation system can learn on line, and the accuracy of short video recommendation is further improved.
In conclusion, the recommendation system can evaluate each short video by combining multiple dimensions, improve the accuracy of short video recommendation, associate the effective play rate target with the two targets of the play completion degree and the screen interaction rate in a product form during training, reduce the influence of an unplayed sample on the play completion degree and the screen interaction rate, and improve the accuracy of short video recommendation.
Continuing with the exemplary structure of the information recommendation device 655 provided by the embodiments of the present invention implemented as software modules, in some embodiments, as shown in fig. 2, the software modules stored in the information recommendation device 655 of the memory 640 may include:
the receiving module 6551 is configured to receive an information browsing request of a target object, and obtain a historical browsing record of the target object according to the information browsing request;
the feature extraction module 6552 is configured to obtain a recommendation information set for the target object according to the historical browsing record; the recommendation information set is a set consisting of recommendation information; obtaining recommendation characteristics corresponding to each recommendation information from each recommendation information in the recommendation information set and the historical browsing records; the recommendation characteristics represent the combined information of each recommendation information and the historical browsing records;
a recommending module 6553, configured to calculate recommendation parameter information for each piece of recommendation information by using a preset recommending model and the recommendation feature; the recommendation parameter information comprises at least one of a playing completion degree, an effective playing rate, a positive feedback rate and the screen interaction rate; the preset recommendation model is a model established by taking the recommendation parameter information as a target; and recommending the recommendation information set based on the recommendation parameter information.
In some embodiments of the present invention, the recommending module 6553 is specifically configured to sort the recommended information set according to the recommended parameter information, so as to obtain a sorting result; and presenting the sequencing result to complete the recommendation of the recommendation information set.
In some embodiments of the present invention, the feature extraction module 6552 is specifically configured to obtain at least one sparse feature and at least one dense feature for each piece of recommendation information and the historical browsing record; the sparse features are high-dimensional features needing dimension reduction, and the dense features are directly obtained features; performing dimensionality reduction on the at least one sparse feature by using at least one preset feature matrix to obtain at least one dimensionality reduction feature; the preset feature matrix is composed of embedded vectors corresponding to the sparse features; and splicing the at least one dimension reduction feature and the at least one dense feature to obtain the recommended feature.
In some embodiments of the present invention, the feature extraction module 6552 is specifically configured to combine each piece of recommendation information with the historical browsing record respectively to obtain combined information corresponding to each piece of recommendation information; extracting the at least one sparse feature from the combined information corresponding to each piece of recommendation information; and extracting the at least one dense feature from the combined information corresponding to each piece of recommendation information.
In some embodiments of the present invention, the feature extraction module 6552 is specifically configured to determine a preset feature matrix for each of the at least one sparse feature respectively; and determining an embedded vector corresponding to each sparse feature according to each sparse feature and the preset feature matrix to obtain the at least one dimension reduction feature.
In some embodiments of the present invention, the recommending module 6553 is specifically configured to perform vector transformation on the recommended features by using a sharing layer of the preset recommending model to obtain an output vector of the sharing layer; calculating the recommendation parameters of the output vector by using each independent layer of the at least one independent layer to obtain the recommendation parameter information; wherein the parameters of each independent layer are independent of each other.
In some embodiments of the present invention, the recommending module 6553 is specifically configured to calculate a recommendation score for each piece of recommendation information according to the recommendation parameter information; the recommendation score represents the interest degree of the target object in each piece of recommendation information; and sorting the recommendation information set by using the recommendation score to obtain the sorting result.
In some embodiments of the present invention, the feature extraction module 6552 is specifically configured to obtain a category tag corresponding to each browsing record in the historical browsing records; counting the common category labels of the target object from the category labels corresponding to each browsing record; and screening a total information set according to the common category label to determine a recommended information set of the target object.
In some embodiments of the present invention, the information recommendation device 655 may further include: a model training module 6553;
the model training module 6553 is specifically configured to construct at least one temporary model by using at least one set of hyper-parameters and the initial recommendation model; training the at least one temporary model by using training sample data to obtain at least one loss function value corresponding to at least one independent layer; the training sample data comprises a training browsing record and a training recommendation information set; acquiring current training parameters respectively corresponding to the at least one temporary model; combining the at least one loss function value and the current training parameters respectively corresponding to the at least one temporary model, and adjusting the current training parameters respectively corresponding to the at least one temporary model according to a combination result to obtain at least one group of training parameters; and forming at least one intermediate recommendation model by using the at least one group of training parameters and the at least one temporary model, and selecting the preset recommendation model from the at least one intermediate recommendation model by using test sample data.
In some embodiments of the present invention, the model training module 6553 is further specifically configured to determine a training recommendation feature for the training sample data; the training sample data is provided with a training label; the training labels comprise at least one of a playing completion label, an effective playing rate label, a front feedback rate label and a screen interaction rate label; predicting the training recommendation characteristics by using at least one temporary model to obtain a prediction result; the prediction result comprises at least one of a prediction playing completion degree, a prediction effective playing rate, a prediction front feedback rate and a prediction screen interaction rate; and obtaining at least one loss function value corresponding to at least one independent layer by using the prediction result and the training label.
In some embodiments of the present invention, the model training module 6553 is further configured to, when the prediction result includes the predicted playing completion, the predicted effective playing rate, the predicted front feedback rate, and the predicted screen interaction rate, predict the training recommendation feature by using the at least one temporary model to obtain a temporary playing completion, the predicted effective playing rate, the predicted front feedback rate, and a temporary screen interaction rate; determining the predicted playing completion degree according to the temporary playing completion degree and the predicted effective playing rate; determining the predicted screen interaction rate according to the predicted effective playing rate and the temporary screen interaction rate; and obtaining the prediction result by utilizing the prediction playing completion degree, the prediction effective playing rate, the prediction front feedback rate and the plane interaction rate.
In some embodiments of the present invention, the model training module 6553 is further configured to perform weighted combination on the current training parameters corresponding to the at least one loss function value and the at least one temporary model, respectively, to obtain a combination result; updating the weight of each independent layer in the at least one independent layer respectively according to the combination result to obtain the updated weight of each independent layer; forming an adjusted current training parameter corresponding to each temporary model in the at least one temporary model by using the updated weight of each independent layer; and combining the adjusted current training parameters corresponding to each temporary model to obtain at least one group of training parameters.
Embodiments of the present invention provide a storage medium storing executable instructions, which, when executed by a processor, cause the processor to perform an information recommendation method provided by embodiments of the present invention, for example, as shown in fig. 3, 6 and 11.
In some embodiments, the storage medium may be memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, 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, the executable information recommendation instructions may be written in any form of programming language (including compiled or interpreted languages), in the form of a program, software module, script, 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, the executable information recommendation instructions may, but need not, 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 in a HyperText Markup Language (HTML) 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, the executable information recommendation instructions may be deployed to be executed on one computing device or on multiple computing devices located 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 (16)

1. An information recommendation method, comprising:
receiving an information browsing request of a target object, and acquiring a historical browsing record of the target object according to the information browsing request;
obtaining a recommendation information set aiming at the target object according to the historical browsing record; the recommendation information set is a set consisting of recommendation information;
obtaining recommendation characteristics corresponding to each recommendation information from each recommendation information in the recommendation information set and the historical browsing records; the recommendation characteristics represent the combined information of each recommendation information and the historical browsing records;
calculating recommendation parameter information for each piece of recommendation information by using a preset recommendation model and the recommendation characteristics; the recommendation parameter information comprises at least one of playing completion degree, effective playing rate, positive feedback rate and screen interaction rate; the preset recommendation model is a model established by taking the recommendation parameter information as a target;
and recommending the recommendation information set based on the recommendation parameter information.
2. The method of claim 1, wherein recommending the set of recommendation information based on the recommendation parameter information comprises:
sorting the recommendation information set according to the recommendation parameter information to obtain a sorting result;
and presenting the sequencing result to complete the recommendation of the recommendation information set.
3. The method according to claim 1, wherein the obtaining of the recommendation characteristic corresponding to each recommendation information from each recommendation information in the recommendation information set and the historical browsing record comprises:
extracting at least one sparse feature and at least one dense feature for each piece of recommendation information and the historical browsing record; the sparse features are high-dimensional features needing dimension reduction, and the dense features are directly obtained features;
performing dimensionality reduction on the at least one sparse feature by using at least one preset feature matrix to obtain at least one dimensionality reduction feature; the preset feature matrix is composed of embedded vectors corresponding to the sparse features;
and splicing the at least one dimension reduction feature and the at least one dense feature to obtain the recommended feature.
4. The method of claim 3, wherein the extracting at least one sparse feature and at least one dense feature for each of the recommendation information and the historical browsing records comprises:
combining each piece of recommendation information with the historical browsing record to obtain combined information corresponding to each piece of recommendation information;
extracting the at least one sparse feature from the combined information corresponding to each piece of recommendation information;
and extracting the at least one dense feature from the combined information corresponding to each piece of recommendation information.
5. The method according to claim 3 or 4, wherein the performing dimension reduction on the at least one sparse feature by using at least one preset feature matrix to obtain at least one dimension-reduced feature comprises:
respectively determining a preset feature matrix for each sparse feature in the at least one sparse feature;
and determining an embedded vector corresponding to each sparse feature according to each sparse feature and the preset feature matrix to obtain the at least one dimension reduction feature.
6. The method of claim 1, wherein the preset recommendation model comprises a shared layer and at least one independent layer; the calculating recommendation parameter information for each piece of recommendation information by using a preset recommendation model and the recommendation characteristics comprises:
utilizing the sharing layer to perform vector conversion on the recommended features to obtain an output vector of the sharing layer;
calculating the recommendation parameters of the output vector by using each independent layer of the at least one independent layer to obtain the recommendation parameter information; wherein the parameters of each independent layer are independent of each other.
7. The method according to claim 2, wherein the sorting the recommendation information sets according to the recommendation parameter information to obtain a sorting result comprises:
calculating a recommendation score for each piece of recommendation information according to the recommendation parameter information; the recommendation score represents the interest degree of the target object in each piece of recommendation information;
and sorting the recommendation information set by using the recommendation score to obtain the sorting result.
8. The method of claim 1, wherein obtaining the set of recommendation information for the target object according to the historical browsing records comprises:
acquiring a category label corresponding to each browsing record in the historical browsing records;
counting the common category labels of the target object from the category labels corresponding to each browsing record;
and screening a total information set according to the common category label to determine a recommended information set of the target object.
9. The method according to claim 1, wherein the recommendation parameter information further includes at least one of a play duration and comment information; the playing duration represents the time length of the target object playing the recommendation information, and the comment information represents the probability of the target object commenting on the recommendation information.
10. The method according to any one of claims 1 to 8, wherein before the receiving an information browsing request of a target object and obtaining a history browsing record of the target object according to the information browsing request, the method further comprises:
constructing at least one temporary model by utilizing at least one group of hyper-parameters and the initial recommendation model;
training the at least one temporary model by using training sample data to obtain at least one loss function value corresponding to at least one independent layer; the training sample data comprises a training browsing record and a training recommendation information set;
acquiring current training parameters respectively corresponding to the at least one temporary model;
combining the at least one loss function value and the current training parameters respectively corresponding to the at least one temporary model, and adjusting the current training parameters respectively corresponding to the at least one temporary model according to a combination result to obtain at least one group of training parameters;
and forming at least one intermediate recommendation model by using the at least one group of training parameters and the at least one temporary model, and selecting the preset recommendation model from the at least one intermediate recommendation model by using test sample data.
11. The method of claim 10, wherein said training said at least one temporary model with training sample data to obtain at least one loss function value for at least one independent layer comprises:
determining training recommendation characteristics aiming at the training sample data; the training sample data is provided with a training label; the training labels comprise at least one of a playing completion label, an effective playing rate label, a front feedback rate label and a screen interaction rate label;
predicting the training recommendation characteristics by using at least one temporary model to obtain a prediction result; the prediction result comprises at least one of a prediction playing completion degree, a prediction effective playing rate, a prediction front feedback rate and a prediction screen interaction rate;
and obtaining at least one loss function value corresponding to at least one independent layer by using the prediction result and the training label.
12. The method of claim 11, wherein when the prediction results include the predicted playing completion, the predicted effective playing rate, the predicted positive feedback rate, and the predicted screen interaction rate, the predicting the training recommendation features using the at least one temporary model to obtain prediction results comprises:
predicting the training recommendation characteristics by using the at least one temporary model to obtain a temporary playing completion degree, the predicted effective playing rate, the predicted front feedback rate and a temporary screen interaction rate;
determining the predicted playing completion degree according to the temporary playing completion degree and the predicted effective playing rate;
determining the predicted screen interaction rate according to the predicted effective playing rate and the temporary screen interaction rate;
and obtaining the prediction result by utilizing the prediction playing completion degree, the prediction effective playing rate, the prediction front feedback rate and the prediction screen interaction rate.
13. The method according to claim 10, wherein said combining the at least one loss function value with the current training parameters respectively corresponding to the at least one temporary model, and adjusting the current training parameters respectively corresponding to the at least one temporary model according to the combination result to obtain the at least one set of training parameters comprises:
performing weighted combination on the current training parameters respectively corresponding to the at least one loss function value and the at least one temporary model to obtain a combined result;
updating the weight of each independent layer in the at least one independent layer respectively according to the combination result to obtain the updated weight of each independent layer;
forming an adjusted current training parameter corresponding to each temporary model in the at least one temporary model by using the updated weight of each independent layer;
and combining the adjusted current training parameters corresponding to each temporary model to obtain at least one group of training parameters.
14. An information recommendation apparatus, comprising:
the receiving module is used for receiving an information browsing request of a target object and acquiring a historical browsing record of the target object according to the information browsing request;
the characteristic extraction module is used for obtaining a recommendation information set aiming at the target object from the historical browsing records; the recommendation information set is a set consisting of recommendation information; obtaining recommendation characteristics corresponding to each recommendation information from each recommendation information in the recommendation information set and the historical browsing records; the recommendation characteristics represent the combined information of each recommendation information and the historical browsing records;
the recommendation module is used for calculating recommendation parameter information aiming at each piece of recommendation information by utilizing a preset recommendation model and the recommendation characteristics; the recommendation parameter information comprises at least one of playing completion degree, effective playing rate, positive feedback rate and screen interaction rate; the preset recommendation model is a model established by taking the recommendation parameter information as a target; and recommending the recommendation information set based on the recommendation parameter information.
15. An information recommendation apparatus characterized by comprising:
a memory for storing executable information recommendation instructions;
a processor for implementing the method of any one of claims 1 to 13 when executing the executable information recommendation instructions stored in the memory.
16. A storage medium having stored thereon executable information recommendation instructions for causing a processor to perform the method of any of claims 1 to 13 when executed.
CN201911005267.9A 2019-10-22 2019-10-22 Information recommendation method, device, equipment and storage medium Active CN110781391B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911005267.9A CN110781391B (en) 2019-10-22 2019-10-22 Information recommendation method, device, equipment and storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911005267.9A CN110781391B (en) 2019-10-22 2019-10-22 Information recommendation method, device, equipment and storage medium

Publications (2)

Publication Number Publication Date
CN110781391A true CN110781391A (en) 2020-02-11
CN110781391B CN110781391B (en) 2023-12-12

Family

ID=69386255

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911005267.9A Active CN110781391B (en) 2019-10-22 2019-10-22 Information recommendation method, device, equipment and storage medium

Country Status (1)

Country Link
CN (1) CN110781391B (en)

Cited By (31)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111291266A (en) * 2020-02-13 2020-06-16 腾讯科技（北京）有限公司 Artificial intelligence based recommendation method and device, electronic equipment and storage medium
CN111311104A (en) * 2020-02-27 2020-06-19 第四范式（北京）技术有限公司 Configuration file recommendation method, device and system
CN111310053A (en) * 2020-03-03 2020-06-19 上海喜马拉雅科技有限公司 Information recommendation method, device, equipment and storage medium
CN111353068A (en) * 2020-02-28 2020-06-30 腾讯音乐娱乐科技（深圳）有限公司 Video recommendation method and device
CN111523028A (en) * 2020-04-16 2020-08-11 广东小天才科技有限公司 Data recommendation method, device, equipment and storage medium
CN111552549A (en) * 2020-04-26 2020-08-18 北京四达时代软件技术股份有限公司 Multitask request dynamic configuration method and system for intelligent equipment
CN111552824A (en) * 2020-04-26 2020-08-18 杭州哔次元科技有限公司 Cartoon recommendation system based on user habits
CN111666522A (en) * 2020-06-10 2020-09-15 创新奇智(上海)科技有限公司 Information processing method, device, equipment and storage medium
CN111737586A (en) * 2020-08-19 2020-10-02 腾讯科技（深圳）有限公司 Information recommendation method, device, equipment and computer readable storage medium
CN111782873A (en) * 2020-06-29 2020-10-16 掌阅科技股份有限公司 Book video-based book recommendation method, electronic device and storage medium
CN111813992A (en) * 2020-07-14 2020-10-23 四川长虹电器股份有限公司 Sorting system and method for movie recommendation candidate set
CN111859149A (en) * 2020-08-03 2020-10-30 腾讯科技（北京）有限公司 Information recommendation method and device, electronic equipment and storage medium
CN111881358A (en) * 2020-07-31 2020-11-03 北京达佳互联信息技术有限公司 Object recommendation system, method and device, electronic equipment and storage medium
CN112035736A (en) * 2020-07-31 2020-12-04 北京达佳互联信息技术有限公司 Information pushing method and device and server
CN112035740A (en) * 2020-08-19 2020-12-04 广州市百果园信息技术有限公司 Project use duration prediction method, device, equipment and storage medium
CN112153423A (en) * 2020-09-15 2020-12-29 济南雪景网络技术有限公司 User-self-driven short video intelligent recommendation method, electronic equipment and storage medium
CN112149839A (en) * 2020-10-09 2020-12-29 北京百度网讯科技有限公司 High-dimensional feature representation learning method, device, equipment and storage medium
CN112749331A (en) * 2020-06-28 2021-05-04 腾讯科技（深圳）有限公司 Data recommendation method and device and computer-readable storage medium
CN112836486A (en) * 2020-12-09 2021-05-25 天津大学 Group hidden-in-field analysis method based on word vectors and Bert
CN112883258A (en) * 2021-01-11 2021-06-01 北京达佳互联信息技术有限公司 Information recommendation method and device, electronic equipment and storage medium
CN113051431A (en) * 2021-06-02 2021-06-29 杭州当贝网络科技有限公司 Video recommendation method fusing user tendency
CN113132803A (en) * 2021-04-23 2021-07-16 Oppo广东移动通信有限公司 Video watching time length prediction method, device, storage medium and terminal
CN113139667A (en) * 2021-05-07 2021-07-20 深圳他米科技有限公司 Hotel room recommendation method, device, equipment and storage medium based on artificial intelligence
CN113157725A (en) * 2021-04-22 2021-07-23 北京小米移动软件有限公司 Recommendation model determining method, device, equipment and storage medium
CN113343024A (en) * 2021-08-04 2021-09-03 北京达佳互联信息技术有限公司 Object recommendation method and device, electronic equipment and storage medium
CN113573106A (en) * 2020-04-28 2021-10-29 北京达佳互联信息技术有限公司 Model updating method and device for multimedia information and server
CN113837844A (en) * 2021-10-19 2021-12-24 北京邮电大学 Multi-cascade downstream enterprise recommendation system and method and storage medium
CN114661949A (en) * 2022-03-31 2022-06-24 南京硕茂电子科技有限公司 Client data analysis system and method under big data situation
CN114880565A (en) * 2022-05-17 2022-08-09 腾讯科技（深圳）有限公司 Information recommendation method, model training method and related device
CN114896061A (en) * 2022-05-07 2022-08-12 百度在线网络技术（北京）有限公司 Training method of computing resource control model, and computing resource control method and device
CN112464100B (en) * 2020-12-14 2023-04-28 未来电视有限公司 Information recommendation model training method, information recommendation method, device and equipment

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103440335A (en) * 2013-09-06 2013-12-11 北京奇虎科技有限公司 Video recommendation method and device
CN105469263A (en) * 2014-09-24 2016-04-06 阿里巴巴集团控股有限公司 Commodity recommendation method and device
WO2017076073A1 (en) * 2015-11-04 2017-05-11 百度在线网络技术（北京）有限公司 Method and apparatus for search and recommendation
CN107729542A (en) * 2017-10-31 2018-02-23 咪咕音乐有限公司 A kind of information methods of marking and device and storage medium
CN108829808A (en) * 2018-06-07 2018-11-16 麒麟合盛网络技术股份有限公司 A kind of page personalized ordering method, apparatus and electronic equipment
CN109446430A (en) * 2018-11-29 2019-03-08 西安电子科技大学 Method, apparatus, computer equipment and the readable storage medium storing program for executing of Products Show
KR20190064042A (en) * 2017-11-30 2019-06-10 삼성에스디에스 주식회사 Method for recommending based on context-awareness and apparatus thereof
WO2019192261A1 (en) * 2018-04-02 2019-10-10 阿里巴巴集团控股有限公司 Payment mode recommendation method and device and equipment

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103440335A (en) * 2013-09-06 2013-12-11 北京奇虎科技有限公司 Video recommendation method and device
CN105469263A (en) * 2014-09-24 2016-04-06 阿里巴巴集团控股有限公司 Commodity recommendation method and device
WO2017076073A1 (en) * 2015-11-04 2017-05-11 百度在线网络技术（北京）有限公司 Method and apparatus for search and recommendation
CN107729542A (en) * 2017-10-31 2018-02-23 咪咕音乐有限公司 A kind of information methods of marking and device and storage medium
KR20190064042A (en) * 2017-11-30 2019-06-10 삼성에스디에스 주식회사 Method for recommending based on context-awareness and apparatus thereof
WO2019192261A1 (en) * 2018-04-02 2019-10-10 阿里巴巴集团控股有限公司 Payment mode recommendation method and device and equipment
CN108829808A (en) * 2018-06-07 2018-11-16 麒麟合盛网络技术股份有限公司 A kind of page personalized ordering method, apparatus and electronic equipment
CN109446430A (en) * 2018-11-29 2019-03-08 西安电子科技大学 Method, apparatus, computer equipment and the readable storage medium storing program for executing of Products Show

Cited By (47)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111291266B (en) * 2020-02-13 2023-03-21 深圳市雅阅科技有限公司 Artificial intelligence based recommendation method and device, electronic equipment and storage medium
CN111291266A (en) * 2020-02-13 2020-06-16 腾讯科技（北京）有限公司 Artificial intelligence based recommendation method and device, electronic equipment and storage medium
CN111311104B (en) * 2020-02-27 2023-08-25 第四范式（北京）技术有限公司 Recommendation method, device and system for configuration file
CN111311104A (en) * 2020-02-27 2020-06-19 第四范式（北京）技术有限公司 Configuration file recommendation method, device and system
CN111353068A (en) * 2020-02-28 2020-06-30 腾讯音乐娱乐科技（深圳）有限公司 Video recommendation method and device
CN111310053A (en) * 2020-03-03 2020-06-19 上海喜马拉雅科技有限公司 Information recommendation method, device, equipment and storage medium
CN111523028A (en) * 2020-04-16 2020-08-11 广东小天才科技有限公司 Data recommendation method, device, equipment and storage medium
CN111552549A (en) * 2020-04-26 2020-08-18 北京四达时代软件技术股份有限公司 Multitask request dynamic configuration method and system for intelligent equipment
CN111552824A (en) * 2020-04-26 2020-08-18 杭州哔次元科技有限公司 Cartoon recommendation system based on user habits
CN111552824B (en) * 2020-04-26 2021-06-04 杭州哔次元科技有限公司 Cartoon recommendation system based on user habits
CN113573106A (en) * 2020-04-28 2021-10-29 北京达佳互联信息技术有限公司 Model updating method and device for multimedia information and server
CN113573106B (en) * 2020-04-28 2023-03-21 北京达佳互联信息技术有限公司 Model updating method and device for multimedia information and server
CN111666522A (en) * 2020-06-10 2020-09-15 创新奇智(上海)科技有限公司 Information processing method, device, equipment and storage medium
CN112749331B (en) * 2020-06-28 2023-09-19 腾讯科技（深圳）有限公司 Data recommendation method and device and computer readable storage medium
CN112749331A (en) * 2020-06-28 2021-05-04 腾讯科技（深圳）有限公司 Data recommendation method and device and computer-readable storage medium
CN111782873A (en) * 2020-06-29 2020-10-16 掌阅科技股份有限公司 Book video-based book recommendation method, electronic device and storage medium
CN111782873B (en) * 2020-06-29 2023-09-08 掌阅科技股份有限公司 Book recommendation method based on book video, electronic equipment and storage medium
CN111813992A (en) * 2020-07-14 2020-10-23 四川长虹电器股份有限公司 Sorting system and method for movie recommendation candidate set
CN111881358B (en) * 2020-07-31 2021-08-03 北京达佳互联信息技术有限公司 Object recommendation system, method and device, electronic equipment and storage medium
CN112035736B (en) * 2020-07-31 2023-12-26 北京达佳互联信息技术有限公司 Information pushing method, device and server
CN111881358A (en) * 2020-07-31 2020-11-03 北京达佳互联信息技术有限公司 Object recommendation system, method and device, electronic equipment and storage medium
CN112035736A (en) * 2020-07-31 2020-12-04 北京达佳互联信息技术有限公司 Information pushing method and device and server
CN111859149A (en) * 2020-08-03 2020-10-30 腾讯科技（北京）有限公司 Information recommendation method and device, electronic equipment and storage medium
CN111737586A (en) * 2020-08-19 2020-10-02 腾讯科技（深圳）有限公司 Information recommendation method, device, equipment and computer readable storage medium
CN112035740A (en) * 2020-08-19 2020-12-04 广州市百果园信息技术有限公司 Project use duration prediction method, device, equipment and storage medium
CN111737586B (en) * 2020-08-19 2020-12-04 腾讯科技（深圳）有限公司 Information recommendation method, device, equipment and computer readable storage medium
CN112035740B (en) * 2020-08-19 2024-06-04 广州市百果园信息技术有限公司 Project use time length prediction method, device, equipment and storage medium
CN112153423A (en) * 2020-09-15 2020-12-29 济南雪景网络技术有限公司 User-self-driven short video intelligent recommendation method, electronic equipment and storage medium
CN112149839A (en) * 2020-10-09 2020-12-29 北京百度网讯科技有限公司 High-dimensional feature representation learning method, device, equipment and storage medium
CN112149839B (en) * 2020-10-09 2024-06-07 北京百度网讯科技有限公司 Method, device, equipment and storage medium for learning representation of high-dimensional characteristics
CN112836486B (en) * 2020-12-09 2022-06-03 天津大学 Group hidden-in-place analysis method based on word vectors and Bert
CN112836486A (en) * 2020-12-09 2021-05-25 天津大学 Group hidden-in-field analysis method based on word vectors and Bert
CN112464100B (en) * 2020-12-14 2023-04-28 未来电视有限公司 Information recommendation model training method, information recommendation method, device and equipment
CN112883258B (en) * 2021-01-11 2024-05-21 北京达佳互联信息技术有限公司 Information recommendation method and device, electronic equipment and storage medium
WO2022148038A1 (en) * 2021-01-11 2022-07-14 北京达佳互联信息技术有限公司 Information recommendation method and device
CN112883258A (en) * 2021-01-11 2021-06-01 北京达佳互联信息技术有限公司 Information recommendation method and device, electronic equipment and storage medium
CN113157725A (en) * 2021-04-22 2021-07-23 北京小米移动软件有限公司 Recommendation model determining method, device, equipment and storage medium
CN113132803A (en) * 2021-04-23 2021-07-16 Oppo广东移动通信有限公司 Video watching time length prediction method, device, storage medium and terminal
CN113139667B (en) * 2021-05-07 2024-02-20 深圳他米科技有限公司 Hotel room recommending method, device, equipment and storage medium based on artificial intelligence
CN113139667A (en) * 2021-05-07 2021-07-20 深圳他米科技有限公司 Hotel room recommendation method, device, equipment and storage medium based on artificial intelligence
CN113051431A (en) * 2021-06-02 2021-06-29 杭州当贝网络科技有限公司 Video recommendation method fusing user tendency
CN113343024A (en) * 2021-08-04 2021-09-03 北京达佳互联信息技术有限公司 Object recommendation method and device, electronic equipment and storage medium
CN113837844A (en) * 2021-10-19 2021-12-24 北京邮电大学 Multi-cascade downstream enterprise recommendation system and method and storage medium
CN114661949B (en) * 2022-03-31 2023-11-24 布谷鸟数字科技(中山)股份有限公司 Client data analysis system and method under big data situation
CN114661949A (en) * 2022-03-31 2022-06-24 南京硕茂电子科技有限公司 Client data analysis system and method under big data situation
CN114896061A (en) * 2022-05-07 2022-08-12 百度在线网络技术（北京）有限公司 Training method of computing resource control model, and computing resource control method and device
CN114880565A (en) * 2022-05-17 2022-08-09 腾讯科技（深圳）有限公司 Information recommendation method, model training method and related device

Also Published As

Publication number Publication date
CN110781391B (en) 2023-12-12

Similar Documents

Publication Publication Date Title
CN110781391B (en) Information recommendation method, device, equipment and storage medium
CN111241311B (en) Media information recommendation method and device, electronic equipment and storage medium
US10958748B2 (en) Resource push method and apparatus
US20220215032A1 (en) Ai-based recommendation method and apparatus, electronic device, and storage medium
CN109543111B (en) Recommendation information screening method and device, storage medium and server
US11232142B2 (en) Flexible real estate search
US11017024B2 (en) Media content rankings for discovery of novel content
CN110781321B (en) Multimedia content recommendation method and device
CN111444428A (en) Information recommendation method and device based on artificial intelligence, electronic equipment and storage medium
US20220284327A1 (en) Resource pushing method and apparatus, device, and storage medium
WO2022016522A1 (en) Recommendation model training method and apparatus, recommendation method and apparatus, and computer-readable medium
CN111090756B (en) Artificial intelligence-based multi-target recommendation model training method and device
US10185983B2 (en) Least-ask: conversational recommender system with minimized user interaction
CN111400603A (en) Information pushing method, device and equipment and computer readable storage medium
WO2019109724A1 (en) Item recommendation method and device
CN111242310A (en) Feature validity evaluation method and device, electronic equipment and storage medium
CN110781377B (en) Article recommendation method and device
CN111258484A (en) Video playing method and device, electronic equipment and storage medium
CN113761359B (en) Data packet recommendation method, device, electronic equipment and storage medium
CN114817692A (en) Method, device and equipment for determining recommended object and computer storage medium
CN113836388A (en) Information recommendation method and device, server and storage medium
CN116777529B (en) Object recommendation method, device, equipment, storage medium and program product
WO2024113641A1 (en) Video recommendation method and apparatus, and electronic device, computer-readable storage medium and computer program product
CN116955842A (en) Content distribution method, apparatus, device and computer readable storage medium
Wang et al. Interactive Art Design and Color Perception Based on the Sparse Network and the VR System

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
REG Reference to a national code

Ref country code: HK

Ref legal event code: DE

Ref document number: 40021443

Country of ref document: HK

SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
TA01 Transfer of patent application right
TA01 Transfer of patent application right

Effective date of registration: 20221121

Address after: 1402, Floor 14, Block A, Haina Baichuan Headquarters Building, No. 6, Baoxing Road, Haibin Community, Xin'an Street, Bao'an District, Shenzhen, Guangdong 518133

Applicant after: Shenzhen Yayue Technology Co.,Ltd.

Address before: 518000 Tencent Building, No. 1 High-tech Zone, Nanshan District, Shenzhen City, Guangdong Province, 35 Floors

Applicant before: TENCENT TECHNOLOGY (SHENZHEN) Co.,Ltd.

GR01 Patent grant
GR01 Patent grant