CN115905690A - Resource recommendation method and device, electronic equipment and storage medium - Google Patents

Resource recommendation method and device, electronic equipment and storage medium Download PDF

Info

Publication number
CN115905690A
CN115905690A CN202211415202.3A CN202211415202A CN115905690A CN 115905690 A CN115905690 A CN 115905690A CN 202211415202 A CN202211415202 A CN 202211415202A CN 115905690 A CN115905690 A CN 115905690A
Authority
CN
China
Prior art keywords
resource
candidate
target
sequence
resources
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.)
Pending
Application number
CN202211415202.3A
Other languages
Chinese (zh)
Inventor
陈志远
潘伟
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.)
Beijing Jingdong Century Trading Co Ltd
Beijing Wodong Tianjun Information Technology Co Ltd
Original Assignee
Beijing Jingdong Century Trading Co Ltd
Beijing Wodong Tianjun Information Technology 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 Beijing Jingdong Century Trading Co Ltd, Beijing Wodong Tianjun Information Technology Co Ltd filed Critical Beijing Jingdong Century Trading Co Ltd
Priority to CN202211415202.3A priority Critical patent/CN115905690A/en
Publication of CN115905690A publication Critical patent/CN115905690A/en
Pending legal-status Critical Current

Links

Images

Landscapes

  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The application provides a resource recommendation method, a resource recommendation device, an electronic device and a storage medium, wherein the method comprises the following steps: generating a plurality of candidate sequences based on first characteristic information related to a user and second characteristic information of a plurality of candidate resources; for at least one target resource in each candidate sequence, determining the estimated exposure probability and the estimated click rate of the target resource based on the corresponding second characteristic information and the corresponding resource display position, and the second characteristic information and the corresponding resource display position of at least one other target resource in the candidate sequence and the first characteristic information; and determining a target sequence from the candidate sequences based on the estimated exposure probability and the estimated click rate, and recommending resources according to the resource display position corresponding to each target resource. A plurality of target resources with higher matching degree with the user can be obtained, the accuracy and effectiveness of the recommended plurality of target resources are improved, and the service quality and the user experience of the network resource platform are improved.

Description

Resource recommendation method and device, electronic equipment and storage medium
Technical Field
The present application relates to the field of information recommendation and artificial intelligence technologies, and in particular, to a resource recommendation method and apparatus, an electronic device, and a storage medium.
Background
With the development of internet technology, network resource platforms such as e-commerce platforms, video playing platforms or news platforms at present all bear massive resources, such as commercial advertisements, articles or videos, and how to select appropriate resources from the massive resources to recommend the resources to users is the core technical capability of the network resource platforms.
In the related technology, for each resource in the mass resources, the click probability of the resource by the user is estimated based on the complex user characteristics and the characteristics of the resource, all the resources are scored based on the click probability of each resource, all the resources are sorted greedily from high to low according to the score, and the resource recommendation is performed according to the sorting result. Because the plurality of resources are mutually influenced when the plurality of resources are recommended, for example, the plurality of resources are alternately displayed in the same area in the page, the click probability of the resource is different for the same resource under the condition that other commonly recommended resources are different. In the related technology, only the characteristics of each resource are considered, and the influence of other resources on the resource is ignored, so that the grading of each resource is not accurate, the matching degree of the whole of a plurality of resources recommended to the user by the network resource platform and the user is low, and correspondingly, the accuracy and the effectiveness of the whole of the recommended plurality of resources are poor, and the service quality and the user experience of the network resource platform are influenced.
Disclosure of Invention
The present application is directed to solving, at least to some extent, one of the technical problems in the related art.
The embodiment of the application provides a resource recommendation method and device, an electronic device and a storage medium, and aims to solve the technical problems that the matching degree of the whole of a plurality of resources recommended to a user by a network resource platform and the user is low, the accuracy and effectiveness of the recommended whole plurality of resources are poor, and the service quality and the user experience of the network resource platform are influenced in the resource recommendation method in the related technology.
An embodiment of a first aspect of the present application provides a resource recommendation method, including: acquiring a resource recommendation request, wherein the request comprises first characteristic information related to a user and the number N of resource display positions, and N is an integer greater than 1; acquiring second characteristic information of a plurality of candidate resources, and generating a plurality of candidate sequences based on the first characteristic information and the second characteristic information; each candidate sequence comprises N target resources which are selected from a plurality of candidate resources and respectively correspond to one resource display position; for at least one target resource in each candidate sequence, determining the estimated exposure probability and the estimated click rate of the target resource based on the corresponding second characteristic information and the corresponding resource display position, and the second characteristic information and the corresponding resource display position of at least one other target resource except the target resource in the candidate sequence and the first characteristic information; and determining a target sequence to be recommended from the candidate sequences based on the estimated exposure probability and the estimated click rate of each target resource in each candidate sequence, and recommending resources according to the resource display position corresponding to each target resource in the target sequence.
An embodiment of a second aspect of the present application provides a resource recommendation apparatus, including: the resource recommendation system comprises a first acquisition module, a second acquisition module and a resource recommendation module, wherein the first acquisition module is used for acquiring a resource recommendation request, the request comprises first characteristic information related to a user and the number N of resource display positions, and N is an integer greater than 1; the generating module is used for acquiring second characteristic information of a plurality of candidate resources and generating a plurality of candidate sequences based on the first characteristic information and the second characteristic information; each candidate sequence comprises N target resources which are selected from a plurality of candidate resources and respectively correspond to one resource display position; the determining module is used for determining the estimated exposure probability and the estimated click rate of the target resource for at least one target resource in each candidate sequence based on the corresponding second characteristic information and the corresponding resource display position, and the second characteristic information and the corresponding resource display position of at least one other target resource except the target resource in the candidate sequence and the first characteristic information; and the recommending module is used for determining a target sequence to be recommended from the candidate sequences based on the estimated exposure probability and the estimated click rate of each target resource in each candidate sequence, and recommending resources according to the resource display position corresponding to each target resource in the target sequence.
An embodiment of a third aspect of the present application provides an electronic device, including: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method for recommending resources as set forth in the embodiments of the first aspect of the present application.
A fourth aspect of the present application is directed to a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the method for recommending resources as set forth in the first aspect of the present application.
An embodiment of the fifth aspect of the present application provides a computer program product, which includes a computer program, and when executed by a processor, the computer program implements the resource recommendation method as set forth in the embodiment of the first aspect of the present application.
One embodiment of the above invention has the following advantages or benefits:
the method comprises the steps of generating a plurality of candidate sequences based on first characteristic information related to a user and second characteristic information of each candidate resource, and determining estimated exposure probability and estimated click rate of the target resource for at least one target resource in the candidate sequences based on the second characteristic information of the target resource and a corresponding resource display position, the second characteristic information of at least one other target resource except the target resource in the candidate sequences and a corresponding resource display position and the first characteristic information, so that for at least one target resource in the candidate sequences, the characteristics of the target resource and the characteristics of other target resources in the candidate sequences can be considered simultaneously, more accurate estimated exposure probability and estimated click rate are obtained, the target sequences are determined from the candidate sequences based on the estimated exposure probability and the estimated click rate, a plurality of target resources with higher matching degree with the user can be obtained, the accuracy and effectiveness of the whole recommended plurality of target resources are improved, and the service quality and user experience of a network resource platform are improved.
Additional aspects and advantages of the present application will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the present application.
Drawings
The foregoing and/or additional aspects and advantages of the present application will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:
fig. 1 is a flowchart illustrating a resource recommendation method according to an embodiment of the present application;
FIG. 2 is a flowchart illustrating a resource recommendation method according to another embodiment of the present application;
FIG. 3 is an architectural diagram of a sequence generation model provided in one embodiment of the present application;
FIG. 4 is a flowchart illustrating a resource recommendation method according to another embodiment of the present application;
FIG. 5 is a diagram illustrating an architecture of an exposure probability prediction model according to an embodiment of the present application;
FIG. 6 is an architecture diagram of a click through rate prediction model according to an embodiment of the present application;
FIG. 7 is a flowchart illustrating a resource recommendation method according to another embodiment of the present application;
FIG. 8 is a schematic structural diagram of an apparatus for recommending resources according to an embodiment of the present application;
FIG. 9 is a schematic structural diagram of an apparatus for recommending resources according to another embodiment of the present application;
FIG. 10 illustrates a block diagram of an exemplary electronic device suitable for use in implementing embodiments of the present application.
Detailed Description
Reference will now be made in detail to embodiments of the present application, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the drawings are exemplary and intended to be used for explaining the present application and should not be construed as limiting the present application.
It should be noted that, in the technical solution of the present application, the acquisition, storage, application, and the like of the personal information of the related user all conform to the regulations of the relevant laws and regulations, and do not violate the customs of the public order.
In the resource recommendation method in the related art, the matching degree of the whole of the plurality of resources recommended to the user by the network resource platform and the user is low, correspondingly, the accuracy and the effectiveness of the whole of the plurality of recommended resources are poor, and the service quality and the user experience of the network resource platform are affected. The method comprises the following steps: acquiring a resource recommendation request, wherein the request comprises first characteristic information related to a user and the number N of resource display positions, and N is an integer greater than 1; acquiring second characteristic information of a plurality of candidate resources, and generating a plurality of candidate sequences based on the first characteristic information and each second characteristic information; each candidate sequence comprises N target resources which are selected from a plurality of candidate resources and correspond to one resource display position; for at least one target resource in each candidate sequence, determining the estimated exposure probability and the estimated click rate of the target resource based on the corresponding second characteristic information and the corresponding resource display position, and the second characteristic information and the corresponding resource display position of at least one other target resource except the target resource in the candidate sequence and the first characteristic information; and determining a target sequence to be recommended from the candidate sequences based on the estimated exposure probability and the estimated click rate of each target resource in each candidate sequence, and recommending the resources according to the resource display position corresponding to each target resource in the target sequence.
Therefore, a plurality of candidate sequences are generated based on first characteristic information related to a user and second characteristic information of each candidate resource, and for at least one target resource in the candidate sequences, estimated exposure probability and estimated click rate of the target resource are determined based on the second characteristic information of the target resource and the corresponding resource display position, the second characteristic information of at least one other target resource except the target resource in the candidate sequences and the corresponding resource display position and the first characteristic information, so that for at least one target resource in the candidate sequences, the characteristics of the target resource and the characteristics of other target resources in the candidate sequences can be considered at the same time, more accurate estimated exposure probability and estimated click rate are obtained, the target sequences are determined from the candidate sequences based on the estimated exposure probability and the estimated click rate, a plurality of target resources with higher overall matching degree with the user can be obtained, the overall accuracy and effectiveness of the recommended plurality of target resources are improved, and the service quality and user experience of a network resource platform are improved.
A resource recommendation method, apparatus, electronic device, storage medium, and computer program product according to embodiments of the present application are described below with reference to the accompanying drawings.
First, a resource recommendation method provided in an embodiment of the present application is described.
Fig. 1 is a flowchart illustrating a resource recommendation method according to an embodiment of the present application. As shown in fig. 1, the resource recommendation method may include the following steps 101-104.
Step 101, a resource recommendation request is obtained, where the request includes first feature information related to a user and the number N of resource display positions, where N is an integer greater than 1.
It should be noted that the resource recommendation method provided in the embodiment of the present application may be executed by a resource recommendation device. The resource recommending device can be an electronic device and can also be configured in the electronic device, so that a plurality of target resources with higher matching degree with the user are obtained by executing the method, the accuracy and effectiveness of the recommended plurality of target resources are improved, and the service quality and the user experience of the network resource platform are improved. The electronic device may be a Personal Computer (PC), a cloud device, a mobile device, a server, and the like, which is not limited in this application. For example, the resource recommendation device may be a network resource platform installed in the electronic device, such as an e-commerce platform, a video playing platform, or a news platform, or a server corresponding to the network resource platform. The embodiment of the present application takes a resource recommendation device as an example of a server corresponding to a network resource platform.
In an embodiment of the application, after a user starts a network resource platform, such as an e-commerce platform, the user firstly enters a home page of the e-commerce platform, and resource recommendation can be performed on the home page, so that a resource recommendation request can be triggered when the network resource platform is started, and correspondingly, a resource recommendation device can obtain the resource recommendation request. Or, when the user switches from other pages of the e-commerce platform to the home page, the switching operation of the user may also trigger the resource recommendation request.
The resource may be a commercial advertisement, an article, a video, etc., which is not limited in this application.
The first feature information may include attribute information of the user, such as an age, a sex, a click behavior in a period of time, whether the user is a high income group, or not. In addition, the first feature information may further include scene information related to a scene in which the user triggers the resource recommendation request, such as time and place when the user triggers the resource recommendation request, and any scene information from which page the user enters a home page of the network resource platform. The present application does not limit the information included in the first feature information. And, the first characteristic information is generally provided by the user and agreed to authorize the related resource recommendation service.
It can be understood that, in order to recommend more resources, the network resource platform generally splits the same area (which may be referred to as an exhibition position) for displaying the resources in the first page or other pages into a plurality of frame bits, each frame bit correspondingly displays one resource, and each frame bit alternately displays the corresponding resources according to a certain sequence. The frame position may be understood as a display position formed by splitting the same region in the page and capable of displaying the corresponding resource in turn.
The resource display position, that is, the position for displaying the resource in the embodiment of the present application may be a certain area for displaying the resource in the top page or other pages of the network resource platform, or may be a certain frame bit in a certain area, which is not limited in the present application. Correspondingly, the N resource display positions may be N regions for displaying resources in a top page or other pages of the network resource platform, or N frame bits of the same region, or other N positions, which is not limited in this application.
The N resource display positions may be sequentially arranged according to a preset arrangement manner, for example, sequentially arranged according to the sequence of display time, or sequentially arranged according to a position relationship, or sequentially arranged according to other arrangement manners, which is not limited in this application.
Step 102, second feature information of a plurality of candidate resources is obtained, and a plurality of candidate sequences are generated based on the first feature information and each second feature information.
Each candidate sequence comprises N target resources which are selected from a plurality of candidate resources and correspond to one resource display position.
The second characteristic information may include attribute information of the candidate resource, such as bid information, a type of the resource to which the candidate resource belongs, an exposure amount in a period of time, a click rate in a period of time, a user who is released, and the like. Wherein, in the case that the candidate resource is a commodity advertisement, the resource type to which the commodity advertisement belongs may include a charging type such as cpc (Cost Per Click, according to a charging mode of deducting fee by clicking), cpm (Cost Per mile, according to a charging mode of deducting fee by exposure), cpd (Cost Per day, according to a charging mode of a daily package), and the like, and commodity types such as a brand, a category, a food type, and the like of the commodity, such as a household appliance type, a clothing type, a food type, and the like; the putting user is the advertiser who puts the commodity advertisement.
In an embodiment of the application, N target resources may be selected from the plurality of candidate resources based on the first feature information and the second feature information, where each target resource corresponds to a resource display position, so that a candidate sequence is generated based on the N target resources. Multiple candidate sequences can be generated by performing multiple selections according to the mode.
Step 103, for at least one target resource in each candidate sequence, determining the estimated exposure probability and the estimated click rate of the target resource based on the corresponding second characteristic information and the corresponding resource display position, and the second characteristic information and the corresponding resource display position of at least one other target resource except the target resource in the candidate sequence and the first characteristic information.
Specifically, after a plurality of candidate sequences are obtained, for each candidate sequence, the estimated exposure probability and the estimated click rate of each target resource can be determined. The pre-estimated exposure probability is the pre-estimated probability of target resource exposure. And the estimated click rate is the ratio of the estimated number of times that the target resource is clicked to the number of times that the target resource is displayed.
In an embodiment of the application, for each target resource in each candidate sequence, the estimated exposure probability and the estimated click rate of the target resource may be determined based on the second feature information corresponding to the target resource, the resource display position corresponding to the target resource, the second feature information of all other target resources except the target resource in the candidate sequence, the resource display positions corresponding to all other target resources, and the first feature information. That is, for each target resource in each candidate sequence, the estimated exposure probability and the estimated click rate of the target resource may be determined based on the second feature information corresponding to all target resources in the candidate sequence, the resource display positions corresponding to all target resources, and the first feature information.
For example, assume that the N resource display locations include four resource display locations a, b, c, and d. A candidate sequence comprises four target resources A, B, C and D, wherein A corresponds to a, B corresponds to B, C corresponds to C, and D corresponds to D. In the embodiment of the application, for any one of the target resources a, B, C, and D, the estimated exposure probability and the estimated click rate of the any one target resource may be determined based on the second feature information of the target resource a, B, C, and D, and the resource display position and the first feature information corresponding to the target resource a, B, C, and D.
In an embodiment of the application, for any one of the partial target resources in each candidate sequence, the estimated exposure probability and the estimated click rate of the target resource may be determined based on the second characteristic information corresponding to the target resource, the resource display position corresponding to the target resource, the second characteristic information of at least one other target resource except the target resource in the candidate sequence, the resource display position corresponding to at least one other target resource, and the first characteristic information; for any one of the other part of the target resources in each candidate sequence, the estimated exposure probability and the estimated click rate of the target resource may be determined based on only the second characteristic information corresponding to the target resource, the resource display position corresponding to the target resource, and the first characteristic information, or a preset value may be determined as the estimated exposure probability and the estimated click rate of the target resource.
Continuing with the above example, assume that a, b, c, and d are 4 frame bits arranged in the display time order, wherein the corresponding resources are displayed cyclically in the order of a, b, c, d, and a. In the embodiment of the application, because a displays the corresponding resource at first, the estimated exposure probability of A can be determined to be 1 for the target resource A corresponding to a, and the estimated click rate of A is determined based on the second characteristic information corresponding to A and the first characteristic information; because B displays the corresponding resource after a, for the target resource B corresponding to B, the estimated exposure probability and the estimated click rate of B can be determined based on the second characteristic information of B, the second characteristic information of the target resource A corresponding to a and the first characteristic information; because C displays the corresponding resources after a and B, for the target resource C corresponding to C, the estimated exposure probability and the estimated click rate of C can be determined based on the second characteristic information of C, the second characteristic information of the target resource A corresponding to a, the second characteristic information of the target resource B corresponding to B and the first characteristic information; because D displays the corresponding resources after a, B and C, for the target resource D corresponding to D, the estimated exposure probability and the estimated click rate of D can be determined based on the second feature information of D, the second feature information of the target resource a corresponding to a, the second feature information of the target resource B corresponding to B, the second feature information of the target resource C corresponding to C and the first feature information.
And 104, determining a target sequence to be recommended from the candidate sequences based on the estimated exposure probability and the estimated click rate of each target resource in each candidate sequence, and recommending resources according to the resource display position corresponding to each target resource in the target sequence.
In an embodiment of the application, for each candidate sequence, the candidate sequence may be scored based on the estimated exposure probability and the estimated click rate of each target resource therein, and then the target sequence to be recommended is determined from the multiple candidate sequences according to the score corresponding to each candidate sequence, so that each target resource in the target sequence may be recommended according to the resource display position corresponding to each target resource in the target sequence.
According to the resource recommendation method provided by the embodiment of the application, a plurality of candidate sequences are generated based on first characteristic information related to a user and second characteristic information of each candidate resource, and for at least one target resource in the candidate sequences, estimated exposure probability and estimated click rate of the target resource are determined based on the second characteristic information of the target resource and a corresponding resource display position, and the second characteristic information of at least one other target resource except the target resource in the candidate sequences and a corresponding resource display position and the first characteristic information, so that for at least one target resource in the candidate sequences, the characteristics of the target resource and the characteristics of other target resources in the candidate sequences can be considered at the same time, more accurate estimated exposure probability and estimated click rate are obtained, the target sequence is determined from the candidate sequences based on the estimated exposure probability and the estimated click rate, a plurality of target resources with higher matching degree with the user can be obtained, the accuracy and effectiveness of the whole recommended plurality of target resources are improved, and the service quality and user experience of a network resource platform are improved.
With reference to fig. 2, a process of generating a plurality of candidate sequences in the resource recommendation method provided in the embodiment of the present application is further described below.
Fig. 2 is a flowchart illustrating a resource recommendation method according to another embodiment of the present application. As shown in fig. 2, the resource recommendation method may include the following steps 201-205.
Step 201, a resource recommendation request is obtained, where the request includes first feature information related to a user and the number N of resource display positions, where N is an integer greater than 1.
The specific implementation process and principle of step 201 may refer to the description of the above embodiments, which is not described herein again.
Step 202, for each candidate resource, obtaining a confidence degree of the candidate resource displayed at each resource display position through a sequence generation model based on the first characteristic information and the plurality of second characteristic information.
The confidence level of a candidate resource displayed at a resource display position represents the probability of the candidate resource being displayed at the resource display position.
It will be appreciated that there are many possibilities to select N candidate resources from a plurality of candidate resources to form a sequence, such as selecting 4 candidate resources from 60 candidate resources to form a sequence, having a 4 60 (about ten million) possibilities. The combined value of each sequence in such a huge fully-arranged sequence space can be estimated, which will seriously exceed the tolerance range of system performance. Therefore, the mapping relation between the candidate resources and the candidate sequences can be modeled through the sequence generation model in the embodiment of the application, so that the performance consumption of the system is reduced.
The sequence generation model is a neural network model used for generating a plurality of candidate sequences.
In one embodiment of the present application, referring to fig. 3, the sequence generation model may include a first embedding layer 301, a first deep neural network layer 302, and a normalization layer 303 connected in sequence. Wherein the first deep neural network layer 302 is a deep neural network considering context information.
Accordingly, step 202 may be implemented by: acquiring a first feature vector corresponding to the first feature information and a second feature vector corresponding to each second feature information based on the first embedding layer 301, and acquiring global feature vectors corresponding to all candidate resources based on each second feature vector; for each candidate resource, splicing the first feature vector, the global feature vector and the corresponding second feature vector to obtain a corresponding first spliced vector; inputting the first stitching vector corresponding to each candidate resource into the first deep neural network layer 302 to obtain a first initial confidence of each candidate resource displayed at each resource display position; each first initial confidence is input into the normalization layer 303 to obtain each confidence corresponding to each candidate resource.
The first characteristic information may include attribute information of the user and scene information related to a scene in which the user triggers the resource recommendation request.
In a possible implementation form, the sequence generation model may further include a feature extraction layer connected to the first embedding layer 301, and the first feature information including the attribute information of the user and the scene information may be input to the feature extraction layer for feature extraction, and the extracted feature vector is input to the first embedding layer 301, so that the first feature vector corresponding to the first feature information is obtained through dimension reduction mapping performed by the first embedding layer 301. Similarly, the second feature information of each candidate resource may be input to the feature extraction layer for feature extraction, and the extracted feature vector is input to the first embedding layer 301, so that the dimension reduction mapping is performed through the first embedding layer 301 to obtain the second feature vector corresponding to the second feature information of each candidate resource.
The sum of the second feature vectors corresponding to the candidate resources may be determined as the global feature vectors corresponding to all the candidate resources, or the second feature vectors corresponding to the candidate resources may be spliced to obtain the global feature vectors corresponding to all the candidate resources, or the global feature vectors corresponding to all the candidate resources may be obtained in other manners, which is not limited in the present application.
Step 203, selecting a plurality of target resources corresponding to each resource display position from the plurality of candidate resources based on the confidence of the display of each candidate resource on each resource display position through a preset selection mode, and generating a plurality of candidate sequences based on the plurality of target resources corresponding to each of the N resource display positions.
Each candidate sequence comprises N target resources which are selected from a plurality of candidate resources and correspond to one resource display position.
The preset selection mode may be a monte carlo sampling mode, an importance sampling mode, or other selection modes, which is not limited in the present application.
In an embodiment of the present application, a target resource corresponding to each of N resource display positions may be selected from a plurality of candidate resources in a monte carlo sampling manner based on a confidence level of each candidate resource displayed at each resource display position, and a candidate sequence may be generated based on one target resource corresponding to each of the N resource display positions. And multiple candidate sequences can be generated by multiple selections in the same way.
The order of each candidate resource is not set, and the second feature information of each candidate resource is directly input into the sequence generation model, so that a plurality of candidate sequences of optimal combinations which are most matched with a user can be determined from the full-array sequence space of the N candidate resources in each candidate resource in a mode of exhaustively exhausting the array of the N candidate resources in each candidate resource, the confidence degree of display of each candidate resource on each resource display position is determined based on the global feature vectors corresponding to all candidate resources in the process of determining the plurality of candidate sequences, the features of other candidate resources can be fully considered for each candidate resource, and the accuracy of the plurality of determined candidate sequences is improved.
It should be noted that the manner of generating the plurality of candidate sequences through the sequence generation model shown in step 202 and step 203 is only an example, and in practical applications, the plurality of candidate sequences may also be generated through other manners, for example, the plurality of candidate sequences may be generated through some heuristic generation strategies, such as a beam search (bundle search) strategy, which is not limited in this application.
And 204, for at least one target resource in each candidate sequence, determining the estimated exposure probability and the estimated click rate of the target resource based on the corresponding second characteristic information and the corresponding resource display position, and the second characteristic information and the corresponding resource display position of at least one other target resource except the target resource in the candidate sequence and the first characteristic information.
Step 205, determining a target sequence to be recommended from the multiple candidate sequences based on the estimated exposure probability and the estimated click rate of each target resource in each candidate sequence, and recommending resources according to the resource display position corresponding to each target resource in the target sequence.
The specific implementation process and principle of steps 204-205 may refer to the description of the above embodiments, and are not described herein again.
In summary, according to the resource recommendation method provided in the embodiment of the application, for each candidate resource, the confidence level of the candidate resource displayed at each resource display position is obtained through a sequence generation model based on the first characteristic information and the plurality of second characteristic information, the plurality of target resources corresponding to each resource display position are selected from the plurality of candidate resources based on the confidence level of each candidate resource displayed at each resource display position through a preset selection mode, and the plurality of candidate sequences are generated based on the plurality of target resources corresponding to the N resource display positions, so that the accuracy of the plurality of determined candidate sequences can be improved. According to at least one target resource in the candidate sequence, the estimated exposure probability and the estimated click rate of the target resource are determined based on the second characteristic information and the corresponding resource display position of the target resource, and the second characteristic information and the corresponding resource display position and the first characteristic information of at least one other target resource except the target resource in the candidate sequence, so that the characteristics of the target resource and the characteristics of other target resources in the candidate sequence can be considered simultaneously for the at least one target resource in the candidate sequence, more accurate estimated exposure probability and estimated click rate are obtained, the target sequence is determined from the candidate sequences based on the estimated exposure probability and the estimated click rate, a plurality of target resources with higher overall matching degree with a user can be obtained, the accuracy and the effectiveness of the overall recommended plurality of target resources are improved, and the service quality and the user experience of a network resource platform are improved.
In a possible implementation form, the N resource display positions may be arranged in sequence, for example, the frame bits may be arranged according to the sequence of the display time, and the corresponding resources are displayed in turn. In addition to the click rate of the resource, whether the resource is exposed or not is also very critical information, and the scoring of each candidate sequence is influenced, so that the accuracy of the finally determined target sequence is influenced, and therefore the exposure probability and the click rate are used as the basis for scoring the candidate sequences in the embodiment of the application. With reference to fig. 4, a process of determining an estimated exposure probability and an estimated click rate of each candidate resource in each candidate sequence in the resource recommendation method provided in the embodiment of the present application is further described below.
Fig. 4 is a flowchart illustrating a resource recommendation method according to another embodiment of the present application. As shown in fig. 4, the resource recommendation method may include the following steps 401-408.
Step 401, a resource recommendation request is obtained, where the request includes first feature information related to a user and the number N of resource display positions, where the N resource display positions are sequentially arranged, and N is an integer greater than 1.
The N resource display positions may be sequentially arranged according to a preset arrangement manner, for example, sequentially arranged according to the sequence of the display time, or sequentially arranged according to the position relationship, or sequentially arranged according to other arrangement manners, which is not limited in the present application.
Step 402, second feature information of a plurality of candidate resources is obtained, and a plurality of candidate sequences are generated based on the first feature information and each second feature information.
Each candidate sequence comprises N target resources which are selected from a plurality of candidate resources and correspond to one resource display position.
The specific implementation process and principle of steps 401 to 402 may refer to the description of the above embodiments, and are not described herein again.
In step 403, for the first resource in each candidate sequence, the corresponding active exposure probability is determined as a first preset value.
It can be understood that, for a scene in which N resource display positions are sequentially arranged, for example, a plurality of frame bits are arranged according to the sequence of display time, and the scene of the corresponding resource is displayed in turn, the N resource display positions may display the corresponding resource in turn based on a preset time interval without user interference; under the condition of user intervention, the corresponding resource can be displayed from a certain resource display position and jumped to other resource display positions to display the corresponding resource based on the behaviors of sliding and the like of the user. Based on this, the estimated exposure probability may be subdivided into an active exposure probability and a passive exposure probability in the embodiments of the present application. The active exposure probability is the exposure probability related to the active behavior of the user, and the passive exposure probability is the exposure probability unrelated to the active behavior of the user. The active behavior of the user may be understood as an operation, such as sliding or clicking, performed actively by the user on a certain resource being displayed.
And the first resource is a target resource corresponding to the resource display position arranged at the head in the candidate sequence. The first preset value may be set as needed, for example, may be set to 1 or other values.
In one embodiment of the present application, the first resource may be set to be necessarily displayed, so that it may be determined that the active exposure probability corresponding to the first resource is 1.
And step 404, for the second resource in each candidate sequence, determining the active exposure probability of the second resource based on the corresponding second characteristic information and the corresponding resource display position, the second characteristic information of the third resource in the candidate sequence and the corresponding resource display position and the first characteristic information through an exposure probability pre-estimation model.
And the second resource is other target resources except the first resource in the candidate sequence. And the third resource is a target resource of which the corresponding resource display position in the candidate sequence is arranged before the second resource.
In an embodiment of the application, since the active exposure probability is related to the active behavior of the user, the active behavior of the user is related to the characteristics of the resource and the characteristics of the user, and for each second resource, a third resource whose corresponding resource display position in the candidate sequence is arranged before the second resource has an influence on the active exposure probability of the second resource, and a target resource whose corresponding resource display position in the candidate sequence is arranged after the second resource has a smaller influence on the active exposure probability of the second resource, for each second resource, the active exposure probability of the second resource may be determined through an exposure probability estimation model based on the second characteristic information corresponding to the second resource, the resource display position corresponding to the second resource, the second characteristic information of the third resource in the candidate sequence, the resource display position corresponding to the third resource, and the first characteristic information.
The exposure probability estimation model is a neural network model used for estimating the estimated exposure probability of each candidate resource in the candidate sequence.
In one embodiment of the present application, referring to fig. 5, the exposure probability prediction model may include a second embedding layer 501, a third embedding layer 502, a stitching layer 503 connected to the second embedding layer 501 and the third embedding layer 502, and a second deep neural network layer 504 connected to the stitching layer 503.
Accordingly, for a second resource in each candidate sequence, the active exposure probability of the second resource may be determined by: acquiring a third feature vector and a fourth feature vector based on the second embedding layer 501, and acquiring a fifth feature vector corresponding to the first feature information based on the third embedding layer 502; inputting the third feature vector, the fourth feature vector, and the fifth feature vector into the stitching layer 503 to obtain a corresponding second stitching vector; the second stitching vector is input into the second deep neural network layer 504 to obtain the active exposure probability P1 of the second resource.
The third feature vector is obtained based on second feature information of the second resource and the corresponding resource display position; the fourth feature vector is obtained based on the second feature information of the third resource and the corresponding resource display position.
In an embodiment of the application, the exposure probability prediction model may further include a feature extraction layer connected to the second embedding layer 501 and the third embedding layer 502, and the feature extraction layer may be input to perform feature extraction on the second feature information of the second resource and the corresponding resource display position, and the second feature information of the third resource and the corresponding resource display position, and input the extracted feature vector to the second embedding layer 501, so that the second embedding layer 501 performs dimension reduction mapping to obtain a third feature vector corresponding to the second resource and a fourth feature vector corresponding to the third resource. In addition, the first feature information including the attribute information of the user and the scene information may be input to the feature extraction layer for feature extraction, and the extracted feature vector may be input to the third embedding layer 502, so that the third embedding layer 502 performs dimension reduction mapping to obtain a fifth feature vector corresponding to the first feature information.
In one embodiment of the present application, steps 403 and 404 may also be replaced by the following manner: for each candidate sequence, acquiring an eighth feature vector corresponding to each target resource based on the second embedding layer 501, and acquiring a fifth feature vector corresponding to the first feature information based on the third embedding layer 502; for each target resource in each candidate sequence, inputting each eighth feature vector and each fifth feature vector into the stitching layer 503, stitching the eighth feature vector corresponding to the target resource, the eighth feature vector corresponding to the other target resource, and the fifth feature vector based on the first weight corresponding to the target resource, the second weight corresponding to the other target resource in the candidate sequence, and the third weight corresponding to the fifth feature vector, to obtain a third stitching vector corresponding to the target resource, and inputting the third stitching vector corresponding to the target resource into the second deep neural network layer 504, to obtain the active exposure probability of the target resource. And obtaining the eighth characteristic vector corresponding to each target resource based on the second characteristic information of the corresponding target resource and the corresponding resource display position.
Therefore, for each target resource in each candidate sequence, the characteristics of other target resources in the candidate sequence can be fully considered, and more accurate active exposure probability is obtained.
In one embodiment of the present application, referring to fig. 5, the exposure probability prediction model may further include a third deep neural network layer 505 connected to the third embedding layer 502. Correspondingly, the passive exposure probability of each target resource in each candidate sequence can be obtained through the following modes: for the first resource in each candidate sequence, determining the corresponding passive exposure probability as a second preset value; and inputting a fifth feature vector into a third deep neural network layer for the second resource in each candidate sequence to obtain a passive exposure probability P2 corresponding to the second resource.
The second preset value can be set as required, for example, can be set to 1 or other values.
In one embodiment of the present application, the first resource may be set to be necessarily displayed, so that the passive exposure probability corresponding to the first resource may be determined to be 1.
In one embodiment of the present application, since the passive exposure probability is independent of the features of the resource, for the second resource in each candidate sequence, the fifth feature vector may be input into the third deep neural network layer, so that the passive exposure probability of the second resource is determined only according to the first feature information of the user.
Therefore, for scenes with N resource display positions arranged in sequence, for example, a plurality of frame positions are arranged according to the sequence of display time, the scenes of corresponding resources are displayed in turn, the estimated exposure probability is subdivided into the active exposure probability and the passive exposure probability, and a multi-task form for simultaneously determining the active exposure probability and the passive exposure probability is adopted for modeling, so that exposure type information can be introduced to enrich the modeling process, the estimated accuracy of the estimated exposure probability of each target resource in a candidate sequence is improved, the accuracy of subsequently scoring each candidate sequence is further improved, and the accuracy of the finally determined target sequence is improved.
Step 405, for each candidate sequence, acquiring a sixth feature vector corresponding to each target resource based on a fourth embedding layer in the click rate estimation model, and acquiring a seventh feature vector corresponding to the first feature information based on a fifth embedding layer in the click rate estimation model.
And obtaining a sixth feature vector corresponding to each target resource based on the corresponding second feature information and the corresponding resource display position.
The click rate estimation model is a list-wise model (a click rate estimation modeling mode) and is a neural network model used for estimating the click rate of each target resource in a candidate sequence, and the input space of the neural network model is the characteristics of the candidate sequence and the characteristics of a user.
In one embodiment of the present application, referring to FIG. 6, the click through rate prediction model may include a fourth embedding layer 601 and a fifth embedding layer 602.
In an embodiment of the present application, the click rate prediction model may further include a feature extraction layer connected to the fourth embedding layer 601, and the second feature information of each target resource in the candidate sequence and the corresponding resource display position may be input to the feature extraction layer for feature extraction, and the extracted feature vector is input to the fourth embedding layer 601, so that dimension reduction mapping is performed through the fourth embedding layer 601, and a sixth feature vector corresponding to each target resource is obtained. In addition, the click rate prediction model may further include a feature extraction layer connected to the fifth embedding layer 602, and the first feature information may be input to the feature extraction layer for feature extraction, and the extracted feature vector is input to the fifth embedding layer 602, so that dimension reduction mapping is performed through the fifth embedding layer 602, and a seventh feature vector corresponding to the first feature information is obtained.
Step 406, inputting each sixth feature vector into a self-attention layer connected to the fourth embedded layer in the click-through rate prediction model, so as to obtain a fusion feature vector corresponding to each target resource based on the self-attention mechanism.
Referring to FIG. 6, the click-through rate prediction model may further include a self-attention layer 603 connected to the fourth embedded layer 601. For at least one target resource in each candidate sequence, the self-attention layer 603 may obtain, based on the self-attention mechanism, a fusion feature vector corresponding to the target resource by combining a plurality of sixth feature vectors.
For example, for each target resource in each candidate sequence, the self-attention layer 603 may determine, based on the self-attention mechanism, a fusion feature vector corresponding to the target resource according to the sixth feature vectors corresponding to all target resources in the sequence and the preset weight corresponding to each target resource. Or, for a first resource in each candidate sequence, the self-attention layer 603 may determine, based on the self-attention mechanism, a fused feature vector corresponding to the first resource according to a sixth feature vector corresponding to the first resource, and for a second resource in each candidate sequence, the self-attention layer 603 may determine, based on the self-attention mechanism, a fused feature vector corresponding to the second resource according to the sixth feature vector corresponding to the second resource and a sixth feature vector corresponding to a third resource whose display position of the corresponding resource in the candidate sequence is arranged before the second resource. Alternatively, the self-attention layer 603 may determine the fused feature vector corresponding to each target resource in each candidate sequence in other manners, which is not limited in this application.
Step 407, inputting each fused feature vector and the seventh feature vector into a fourth deep neural network layer connected with the self-attention layer and the fifth embedded layer in the click rate estimation model to obtain the estimated click rate of each target resource in the candidate sequence.
Referring to fig. 6, a fourth deep neural network layer 604 connected to the self-attention layer 603 and the fifth embedding layer 602 may be further included in the click-through rate prediction model. Taking N as 4 as an example, inputting 4 fusion feature vectors and the seventh feature vector into the fourth deep neural network layer 604, the estimated click rate pctr of 4 target resources in the candidate sequence can be obtained 1 、pctr 2 、pctr 3 、pctr 4 。
As for at least one target resource in the candidate sequence, the self-attention layer in the click rate prediction model can combine with the plurality of sixth feature vectors to obtain the fusion feature vector corresponding to the target resource, and then the prediction click rate of each target resource is determined based on the fusion feature vector corresponding to each target resource, so that the characteristics of other target resources in the candidate sequence can be fully considered for at least one target resource in the candidate sequence, and more accurate click rate is obtained.
In addition, in consideration of certain differences and connections of resources belonging to different resource types in an actual scene, for example, a product advertisement of a cpc charging type is generally a specific product, and a product advertisement of a cpm charging type is generally a product aggregation page and a movable page.
And 408, determining a target sequence to be recommended from the candidate sequences based on the estimated exposure probability and the estimated click rate of each target resource in each candidate sequence, and recommending resources according to the resource display position corresponding to each target resource in the target sequence.
According to the resource recommendation method provided by the embodiment of the application, a plurality of candidate sequences are generated based on first characteristic information related to a user and second characteristic information of each candidate resource, and for at least one target resource in the candidate sequences, estimated exposure probability and estimated click rate of the target resource are determined based on the second characteristic information of the target resource and a corresponding resource display position, and the second characteristic information of at least one other target resource except the target resource in the candidate sequences and a corresponding resource display position and the first characteristic information, so that for at least one target resource in the candidate sequences, the characteristics of the target resource and the characteristics of other target resources in the candidate sequences can be considered at the same time, more accurate estimated exposure probability and estimated click rate are obtained, the target sequence is determined from the candidate sequences based on the estimated exposure probability and the estimated click rate, a plurality of target resources with higher matching degree with the user can be obtained, the accuracy and effectiveness of the whole recommended plurality of target resources are improved, and the service quality and user experience of a network resource platform are improved.
With reference to fig. 7, a process of determining a target sequence to be recommended from a plurality of candidate sequences in the resource recommendation method provided in the embodiment of the present application is further described.
Fig. 7 is a flowchart illustrating a resource recommendation method according to another embodiment of the present application. As shown in fig. 7, the resource recommendation method may include the following steps 701-707.
Step 701, a resource recommendation request is obtained, where the request includes first feature information related to a user and the number N of resource display positions, where N is an integer greater than 1.
Step 702, obtaining second feature information of a plurality of candidate resources, and generating a plurality of candidate sequences based on the first feature information and each second feature information.
Each candidate sequence comprises N target resources which are selected from a plurality of candidate resources and correspond to one resource display position.
Step 703, for at least one target resource in each candidate sequence, determining an estimated exposure probability and an estimated click rate of the target resource based on the corresponding second feature information and the corresponding resource display position, and the second feature information and the corresponding resource display position of at least one other target resource except the target resource in the candidate sequence and the first feature information.
The specific implementation process and principle of steps 701-703 may refer to the description of the above embodiments, and are not described herein again.
And 704, for each candidate sequence, determining the similarity between the target resources based on the second characteristic information of the target resources.
Step 705, determining a score corresponding to the candidate sequence by a preset method in combination with the estimated exposure probability, the estimated click rate, the bid information, the type of the resource to which the candidate sequence belongs, and the similarity between the target resources.
The second characteristic information of each target resource may include bid information of the corresponding target resource and a resource type to which the target resource belongs.
In one embodiment of the present application, for each candidate sequence, the score corresponding to the candidate sequence may be determined by the following formula:
Figure BDA0003939109570000141
wherein pos represents a resource display location; t represents the resource type of the target resource in the candidate sequence; bid price Bid information representing a target resource in the candidate sequence; pctr represents the estimated click rate of a certain target resource in the candidate sequence, alpha t Represents the power of one; boost t The weight corresponding to the t resource type can be set according to needs, and the control of resource recommendation of different resource types can be easily realized by setting the weight; imp _ prob pos Representing the estimated exposure probability of a certain target resource in the candidate sequence, including the sum of the active exposure probability and the passive exposure probability; w is a group of ecpm And W pctr Respectively representing the respective weights of the two targets of consumption and click (namely the weights corresponding to the first two items in the three items included in the formula), wherein the two weights are the pareto optimal weights obtained by the pareto optimizer; sim (ad) pos_i ,ad pos_j ) And representing the similarity between two target resources with pos _ i and pos _ j as the display positions of the corresponding resources in the candidate sequence.
Step 706, determining a target sequence from the plurality of candidate sequences based on the scores corresponding to the plurality of candidate sequences.
Wherein, the candidate sequence with the highest score corresponding to the multiple candidate sequences can be determined as the target sequence.
It can be understood that when multiple target resources that are too similar exist in a candidate sequence, such multiple resources, if recommended to a user, may seriously affect the browsing experience of the user. In the embodiment of the application, the score corresponding to the candidate sequence is determined by adopting the mode shown in the formula, and the penalty on the score can be performed on the candidate sequence with a plurality of over-similar target resources, so that the possibility that the over-similar target resources exist in the finally determined target sequence is reduced, and the browsing experience of a user is improved. Compared with a duplication removing mode based on a preset rule in the related technology, for example, when two adjacent commodity advertisements belong to the same category or the corresponding commodity names of the two commodity advertisements are the same, one commodity advertisement is removed, the method in the embodiment of the application is more flexible, the problem that the duplication removing rule is too strict for a certain user and the duplication removing rule is too loose for another user does not exist, the matching degree of a target sequence and the user can be greatly improved, and the user experience is improved. In addition, by adopting the above mode, the candidate sequence is scored based on the estimated exposure probability, the estimated click rate, the offer information, the type of the belonged resource and the similarity between the target resources in the candidate sequence, the income of the resource recommended by the network resource platform can be improved, the browsing experience of the user and the income of the network resource platform are considered while the participation of the user is considered, the balance between the user experience and the income is realized, and the traffic distribution efficiency of the network resource platform is improved.
And 707, recommending the resources according to the resource display positions corresponding to the target resources in the target sequence.
In an embodiment of the present application, in the manner in the foregoing embodiment, a plurality of candidate sequences are generated based on a sequence generation model, an estimated click rate of each target resource in the candidate sequences is determined based on a click rate estimation model, and when an estimated exposure probability of each target resource in the candidate sequences is determined based on an exposure probability estimation model, the sequence generation model, the click rate estimation model, and the exposure probability estimation model may be obtained through training in the following manner:
acquiring first sample characteristic information and a plurality of sample sequences which are respectively related to a plurality of sample users, wherein each sample sequence comprises N sample resources which are obtained by random sampling and respectively correspond to one resource display position, and acquiring second sample characteristic information of each sample resource; marking each sample resource by the corresponding sample exposure probability and sample sampling rate;
respectively training an initial click rate estimation model and an initial exposure probability estimation model based on the first sample characteristic information, the second sample characteristic information of the sample resources, the corresponding resource display positions, the corresponding sample exposure probability and the sample sampling rate to obtain a trained click rate estimation model and a trained exposure probability estimation model;
and training the initial sequence generation model based on the first sample characteristic information and the second sample characteristic information and by combining the trained click rate estimation model and the trained exposure probability estimation model to obtain the trained sequence generation model.
The first sample characteristic information may include attribute information of the sample user, such as any attribute information of the sample user, such as the age and gender of the sample user, click behavior within a period of time, whether the sample user is a high income group, and the like, and may include scene information related to a scene in which the sample user triggers the resource recommendation request, such as any scene information of the time and place in which the sample user triggers the resource recommendation request, and a first page from which the sample user enters the network resource platform.
The second sample characteristic information may include attribute information of the sample resource, such as bid information, the type of the resource to which the sample resource belongs, exposure amount in a period of time, click amount in a period of time, and the like.
Specifically, the sample exposure probability and the sample sampling rate of each sample resource in each sample sequence may be supervised, and the initial click rate estimation model and the initial exposure probability estimation model are trained respectively based on the plurality of first sample characteristic information, the plurality of second sample characteristic information, and the corresponding resource display positions, so as to obtain a trained click rate estimation model and a trained exposure probability estimation model. And training an initial sequence generation model based on the first sample characteristic information and the second sample characteristic information, wherein when the initial sequence generation model is trained, a click rate pre-estimated model after training can be adopted to determine the pre-estimated click rate of each resource in the sequence generated by the initial sequence generation model, a pre-estimated exposure probability of each resource in the sequence generated by the initial sequence generation model is determined by adopting the exposure probability pre-estimated model after training, the sequence generated by the initial sequence generation model is scored based on the pre-estimated click rate and the pre-estimated exposure probability, the offer information of each resource in the sequence generated by the initial sequence generation model, the type of the affiliated resource and the similarity between every two, the scoring is used as supervision, and the initial sequence generation model is trained.
By training the sequence generation model, the exposure probability pre-estimation model and the click rate pre-estimation model in the mode, the prediction results of the trained exposure probability pre-estimation model and click rate pre-estimation model can be fed back to the sequence generation model, so that the possibility that too similar resources exist in the sequence generated by the sequence generation model is reduced, the sequence generated by the sequence generation model is matched with a user better, and the prediction accuracy of the sequence generation model is improved.
Fig. 8 is a schematic structural diagram of a resource recommendation device according to an embodiment of the present application.
As shown in fig. 8, the resource recommending apparatus 800 may include: a first acquisition module 801, a generation module 802, a determination module 803, and a recommendation module 804.
It should be noted that the resource recommendation device provided in this embodiment of the present application may execute the resource recommendation method in the foregoing embodiment, and the resource recommendation device may be an electronic device, and may also be configured in the electronic device, so as to obtain a plurality of target resources with higher overall matching degree with the user by executing the method, improve the overall accuracy and effectiveness of the recommended plurality of target resources, and improve the service quality and the user experience of the network resource platform. The electronic device may be a PC, a cloud device, a mobile device, a server, and the like, which is not limited in this application.
The first obtaining module 801 is configured to obtain a resource recommendation request, where the request includes first feature information related to a user and a number N of resource display positions, where N is an integer greater than 1;
a generating module 802, configured to obtain second feature information of multiple candidate resources, and generate multiple candidate sequences based on the first feature information and each second feature information; each candidate sequence comprises N target resources which are selected from a plurality of candidate resources and correspond to one resource display position;
a determining module 803, configured to determine, for at least one target resource in each candidate sequence, an estimated exposure probability and an estimated click rate of the target resource based on the corresponding second feature information and the corresponding resource display position, and the second feature information and the corresponding resource display position of at least one other target resource except the target resource in the candidate sequence and the first feature information;
and the recommending module 804 is configured to determine a target sequence to be recommended from the multiple candidate sequences based on the estimated exposure probability and the estimated click rate of each target resource in each candidate sequence, and recommend resources according to the resource display position corresponding to each target resource in the target sequence.
It should be noted that the explanation in the foregoing resource recommendation method embodiment is also applicable to the resource recommendation apparatus in this embodiment, and details are not described here.
According to the resource recommending device, a plurality of candidate sequences are generated based on first characteristic information related to a user and second characteristic information of each candidate resource, and for at least one target resource in the candidate sequences, the estimated exposure probability and the estimated click rate of the target resource are determined based on the second characteristic information of the target resource and the corresponding resource display position, and the second characteristic information of at least one other target resource except the target resource in the candidate sequences and the corresponding resource display position and the first characteristic information, so that for at least one target resource in the candidate sequences, the characteristics of the target resource and the characteristics of other target resources in the candidate sequences can be considered at the same time, more accurate estimated exposure probability and estimated click rate are obtained, the target sequences are determined from the candidate sequences based on the estimated exposure probability and the estimated click rate, a plurality of target resources with higher overall matching degree with the user can be obtained, the overall accuracy and effectiveness of the recommended plurality of target resources are improved, and the service quality and user experience of a network resource platform are improved.
Fig. 9 is a schematic structural diagram of a resource recommendation device according to another embodiment of the present application.
As shown in fig. 9, the resource recommending apparatus 900 may include: a first obtaining module 901, a generating module 902, a determining module 903 and a recommending module 904. The first obtaining module 901, the generating module 902, the determining module 903 and the recommending module 904 in fig. 9 have the same functions and structures as the first obtaining module 801, the generating module 802, the determining module 803 and the recommending module 804 in fig. 8.
It should be noted that the resource recommendation device 900 provided in this embodiment of the present application may execute the resource recommendation method in the foregoing embodiment, and the resource recommendation device may be an electronic device, and may also be configured in the electronic device, so as to obtain a plurality of target resources with higher matching degree with the user as a whole by executing the method, improve the accuracy and effectiveness of the recommended plurality of target resources as a whole, and improve the service quality and the user experience of the network resource platform. The electronic device may be a PC, a cloud device, a mobile device, a server, and the like, which is not limited in this application.
In a possible implementation manner of the embodiment of the present application, the generating module 902 includes:
a first obtaining unit 9021, configured to obtain, for each candidate resource, a confidence level of display of the candidate resource at each resource display position based on the first feature information and the plurality of second feature information through a sequence generation model;
the generating unit 9022 is configured to select, in a preset selection manner, multiple target resources corresponding to each resource display position from the multiple candidate resources based on the confidence level of each candidate resource displayed at each resource display position, and generate multiple candidate sequences based on multiple target resources corresponding to each of the N resource display positions.
In another possible implementation manner of the embodiment of the application, the sequence generation model includes a first embedding layer, a first deep neural network layer, and a normalization layer, which are connected in sequence; a first obtaining unit 9021, configured to:
acquiring a first feature vector corresponding to the first feature information and a second feature vector corresponding to each second feature information based on the first embedded layer, and acquiring global feature vectors corresponding to all candidate resources based on each second feature vector;
for each candidate resource, splicing the first feature vector, the global feature vector and the corresponding second feature vector to obtain a corresponding first spliced vector;
inputting the first splicing vector corresponding to each candidate resource into the first deep neural network layer to obtain a first initial confidence coefficient of each candidate resource displayed at each resource display position;
and inputting each first initial confidence coefficient into a normalization layer to obtain each confidence coefficient corresponding to each candidate resource.
In another possible implementation manner of the embodiment of the application, the estimated exposure probability comprises an active exposure probability related to an active behavior of a user, and the N resource display positions are sequentially arranged; a determining module 903, comprising:
a first determining unit 9031, configured to determine, for a first resource in each candidate sequence, a corresponding active exposure probability as a first preset value; the first resource is a target resource corresponding to the resource display position arranged at the head in the candidate sequence;
a second determining unit 9032, configured to determine, through an exposure probability prediction model, an active exposure probability of a second resource based on corresponding second feature information and a corresponding resource display position, and second feature information of a third resource in each candidate sequence and a corresponding resource display position and first feature information, for the second resource in each candidate sequence; the second resource is other target resources except the first resource in the candidate sequence; and the third resource is a target resource of which the corresponding resource display position in the candidate sequence is arranged before the second resource.
In another possible implementation manner of the embodiment of the application, the exposure probability pre-estimation model includes a second embedded layer, a third embedded layer, a splicing layer connected to the second embedded layer and the third embedded layer, and a second deep neural network layer connected to the splicing layer;
a second determining unit 9032, configured to:
acquiring a third feature vector and a fourth feature vector based on the second embedded layer, and acquiring a fifth feature vector corresponding to the first feature information based on the third embedded layer; the third feature vector is obtained based on second feature information of the second resource and the corresponding resource display position; the fourth feature vector is obtained based on second feature information of the third resource and the corresponding resource display position;
inputting the third feature vector, the fourth feature vector and the fifth feature vector into a splicing layer to obtain a corresponding second splicing vector;
and inputting the second splicing vector into a second deep neural network layer to obtain the active exposure probability of the second resource.
In another possible implementation manner of the embodiment of the application, the pre-estimated exposure probability further includes a passive exposure probability unrelated to the active behavior of the user; the exposure probability pre-estimation model further comprises a third deep neural network layer connected with the third embedded layer;
the determining module 903 further includes:
a third determining unit 9033, configured to determine, for the first resource in each candidate sequence, the corresponding passive exposure probability as a second preset value;
a second obtaining unit 9034, configured to, for the second resource in each candidate sequence, input the fifth feature vector into the third deep neural network layer, so as to obtain a corresponding passive exposure probability.
In another possible implementation manner of the embodiment of the present application, the determining module 903 includes:
a third obtaining unit 9035, configured to obtain, for each candidate sequence, a sixth feature vector corresponding to each target resource based on a fourth embedded layer in the click rate prediction model, and obtain a seventh feature vector corresponding to the first feature information based on a fifth embedded layer in the click rate prediction model; the sixth feature vector corresponding to each target resource is obtained based on the corresponding second feature information and the corresponding resource display position;
a fourth obtaining unit 9036, configured to input each sixth feature vector into a self-attention layer connected to the fourth embedded layer in the click rate prediction model, so as to obtain a fusion feature vector corresponding to each target resource based on a self-attention mechanism;
a fifth obtaining unit 9037, configured to input each fused feature vector and the seventh feature vector into a fourth deep neural network layer connected to the self-attention layer and the fifth embedded layer in the click rate prediction model, so as to obtain a predicted click rate of each target resource in the candidate sequence.
In another possible implementation manner of the embodiment of the present application, the second feature information includes bid information of the corresponding target resource and a resource type to which the target resource belongs; a recommendation module 904 comprising:
a fourth determining unit, configured to determine, for each candidate sequence, a similarity between the target resources based on the second feature information of the target resources;
the fifth determining unit is used for determining a score corresponding to the candidate sequence by combining the estimated exposure probability, the estimated click rate, the bid information, the type of the affiliated resource and the similarity between the target resources in a preset mode;
and a sixth determining unit, configured to determine the target sequence from the multiple candidate sequences based on scores corresponding to the multiple candidate sequences.
In another possible implementation manner of the embodiment of the present application, the apparatus 900 for recommending a resource further includes:
the second acquisition module is used for acquiring first sample characteristic information and a plurality of sample sequences which are respectively related to a plurality of sample users, wherein each sample sequence comprises N sample resources which are obtained by random sampling and respectively correspond to one resource display position, and acquiring second sample characteristic information of each sample resource; marking each sample resource by the corresponding sample exposure probability and sample sampling rate;
the first training module is used for respectively training an initial click rate estimation model and an initial exposure probability estimation model based on a plurality of first sample characteristic information, a plurality of second sample characteristic information of sample resources, corresponding resource display positions, corresponding sample exposure probabilities and sample sampling rates so as to obtain a trained click rate estimation model and a trained exposure probability estimation model;
and the second training module is used for training the initial sequence generation model based on the first sample characteristic information and the second sample characteristic information and by combining the trained click rate estimation model and the trained exposure probability estimation model to obtain the trained sequence generation model.
According to the resource recommending device, a plurality of candidate sequences are generated based on first characteristic information related to a user and second characteristic information of each candidate resource, and for at least one target resource in the candidate sequences, the estimated exposure probability and the estimated click rate of the target resource are determined based on the second characteristic information of the target resource and the corresponding resource display position, and the second characteristic information of at least one other target resource except the target resource in the candidate sequences and the corresponding resource display position and the first characteristic information, so that for at least one target resource in the candidate sequences, the characteristics of the target resource and the characteristics of other target resources in the candidate sequences can be considered at the same time, more accurate estimated exposure probability and estimated click rate are obtained, the target sequences are determined from the candidate sequences based on the estimated exposure probability and the estimated click rate, a plurality of target resources with higher overall matching degree with the user can be obtained, the overall accuracy and effectiveness of the recommended plurality of target resources are improved, and the service quality and user experience of a network resource platform are improved.
In order to implement the above embodiments, the present application also provides an electronic device, including: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of recommending resources as set forth in any of the embodiments of the application.
The electronic device may be a Personal Computer (PC), a cloud device, a mobile device, a server, and the like, and the mobile device may be any hardware device such as a mobile phone, a tablet Computer, a Personal digital assistant, a wearable device, and a vehicle-mounted device, which is not limited in this application.
In order to achieve the above embodiments, the present application also proposes a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the method for recommending a resource as proposed in any of the foregoing embodiments of the present application.
In order to implement the above embodiments, the present application also proposes a computer program product comprising a computer program which, when executed by a processor, implements the method for recommending resources as proposed in any of the previous embodiments of the present application.
FIG. 10 shows a schematic block diagram of an example electronic device 1000 that may be used to implement embodiments of the present application. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the present application that are described and/or claimed herein.
As shown in fig. 10, the electronic device 1000 may include a computing unit 1001 that may perform various appropriate actions and processes according to a computer program stored in a Read Only Memory (ROM) 1002 or a computer program loaded from a storage unit 1008 into a Random Access Memory (RAM) 1003. In the RAM1003, various programs and data necessary for the operation of the device 1000 can be stored. The calculation unit 1001, ROM 1002, and RAM1003 are connected to each other by a bus 1004. An input/output (I/O) interface 1005 is also connected to bus 1004.
A number of components in device 1000 are connected to I/O interface 1005, including: an input unit 1006 such as a keyboard, a mouse, and the like; an output unit 1007 such as various types of displays, speakers, and the like; a storage unit 1008 such as a magnetic disk, optical disk, or the like; and a communication unit 1009 such as a network card, a modem, a wireless communication transceiver, or the like. The communication unit 1009 allows the device 1000 to exchange information/data with other devices through a computer network such as the internet and/or various telecommunication networks.
Computing unit 1001 may be a variety of general and/or special purpose processing components with processing and computing capabilities. Some examples of the computing unit 1001 include, but are not limited to, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), various dedicated Artificial Intelligence (AI) computing chips, various computing units running machine learning model algorithms, a Digital Signal Processor (DSP), and any suitable processor, controller, microcontroller, and so forth. The calculation unit 1001 executes the respective methods and processes described above, such as a recommendation method of resources. For example, in some embodiments, the resource recommendation method may be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 1008. In some embodiments, part or all of the computer program may be loaded and/or installed onto device 1000 via ROM 1002 and/or communications unit 1009. When the computer program is loaded into the RAM1003 and executed by the computing unit 1001, one or more steps of the method for recommending resources described above may be performed. Alternatively, in other embodiments, the computing unit 1001 may be configured to perform the recommendation method for resources in any other suitable way (e.g., by means of firmware).
Various implementations of the systems and techniques described here above may be implemented in digital electronic circuitry, integrated circuitry, field Programmable Gate Arrays (FPGAs), application Specific Integrated Circuits (ASICs), application Specific Standard Products (ASSPs), system on a chip (SOCs), complex Programmable Logic Devices (CPLDs), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, receiving data and instructions from, and transmitting data and instructions to, a storage system, at least one input device, and at least one output device.
Program code for implementing the methods of the present application may be written in any combination of one or more programming languages. These program code may be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the program code, when executed by the processor or controller, causes the functions/acts specified in the flowchart and/or block diagram to be performed. The program code may execute entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine or entirely on the remote machine or server.
In the context of this application, a machine-readable medium may be a tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.
To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which a user can provide input to the computer. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic, speech, or tactile input.
The systems and techniques described here can be implemented in a computing system that includes a back-end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), wide Area Networks (WANs), the Internet, and blockchain networks.
The computer system may include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The Server can be a cloud Server, also called a cloud computing Server or a cloud host, and is a host product in a cloud computing service system, so as to solve the defects of high management difficulty and weak service expansibility in the traditional physical host and VPS service ("Virtual Private Server", or simply "VPS"). The server may be a cloud server, a server of a distributed system, or a server with a combined blockchain.
It should be understood that various forms of the flows shown above, reordering, adding or deleting steps, may be used. For example, the steps described in the present disclosure may be executed in parallel, sequentially, or in different orders, and are not limited herein as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved.
The above-described embodiments should not be construed as limiting the scope of the present application. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may be made, depending on design requirements and other factors. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims (20)

1. A method for recommending resources, the method comprising:
acquiring a resource recommendation request, wherein the request comprises first characteristic information related to a user and the number N of resource display positions, and N is an integer greater than 1;
acquiring second characteristic information of a plurality of candidate resources, and generating a plurality of candidate sequences based on the first characteristic information and the second characteristic information; each candidate sequence comprises N target resources which are selected from a plurality of candidate resources and respectively correspond to one resource display position;
for at least one target resource in each candidate sequence, determining an estimated exposure probability and an estimated click rate of the target resource based on corresponding second characteristic information and a corresponding resource display position, and second characteristic information and a corresponding resource display position of at least one other target resource except the target resource in the candidate sequence and the first characteristic information;
and determining a target sequence to be recommended from the candidate sequences based on the estimated exposure probability and the estimated click rate of each target resource in each candidate sequence, and recommending resources according to the resource display position corresponding to each target resource in the target sequence.
2. The method of claim 1, wherein generating a plurality of candidate sequences based on the first feature information and the second feature information comprises:
for each candidate resource, acquiring a confidence degree of the candidate resource displayed at each resource display position through a sequence generation model based on the first characteristic information and a plurality of second characteristic information;
and selecting a plurality of target resources corresponding to each resource display position from the plurality of candidate resources based on the confidence degree of each candidate resource displayed at each resource display position through a preset selection mode, and generating a plurality of candidate sequences based on the plurality of target resources corresponding to the N resource display positions.
3. The method of claim 2, wherein the sequence generative model comprises a first embedding layer, a first deep neural network layer, and a normalization layer connected in sequence; the obtaining, by the sequence generation model, a confidence level of the candidate resource displayed at each resource display position based on the first feature information and the plurality of second feature information includes:
acquiring a first feature vector corresponding to the first feature information and a second feature vector corresponding to each piece of second feature information based on the first embedding layer, and acquiring global feature vectors corresponding to all the candidate resources based on each second feature vector;
for each candidate resource, splicing the first feature vector, the global feature vector and the corresponding second feature vector to obtain a corresponding first spliced vector;
inputting the first stitching vector corresponding to each candidate resource into the first deep neural network layer to obtain a first initial confidence of each candidate resource displayed at each resource display position;
and inputting each first initial confidence degree into the normalization layer to obtain each confidence degree corresponding to each candidate resource.
4. The method according to claim 2 or 3, wherein the estimated exposure probability comprises an active exposure probability related to active actions of a user, and the N resource display positions are arranged in sequence;
the determining the estimated exposure probability and the estimated click rate of the target resource comprises the following steps:
for the first resource in each candidate sequence, determining the corresponding active exposure probability as a first preset value; the first resource is a target resource corresponding to a resource display position arranged at the head in the candidate sequence;
for a second resource in each candidate sequence, determining the active exposure probability of the second resource based on corresponding second characteristic information and a corresponding resource display position, and second characteristic information and a corresponding resource display position of a third resource in the candidate sequence and the first characteristic information through an exposure probability pre-estimation model; the second resource is other target resources except the first resource in the candidate sequence; the third resource is a target resource of which the corresponding resource display position in the candidate sequence is arranged before the second resource.
5. The method of claim 4, wherein the pre-estimated exposure probability model comprises a second embedding layer, a third embedding layer, a splicing layer connected to the second embedding layer and the third embedding layer, and a second deep neural network layer connected to the splicing layer;
the determining the active exposure probability for the second resource comprises:
acquiring a third feature vector and a fourth feature vector based on the second embedded layer, and acquiring a fifth feature vector corresponding to the first feature information based on the third embedded layer; the third feature vector is obtained based on second feature information of the second resource and a corresponding resource display position; the fourth feature vector is obtained based on second feature information of the third resource and a corresponding resource display position;
inputting the third feature vector, the fourth feature vector and the fifth feature vector into the splicing layer to obtain a corresponding second splicing vector;
inputting the second stitching vector into the second deep neural network layer to obtain the active exposure probability of the second resource.
6. The method of claim 5, wherein the estimated exposure probabilities further comprise passive exposure probabilities that are independent of the user's active behavior; the exposure probability pre-estimation model further comprises a third deep neural network layer connected with the third embedded layer;
the method further comprises the following steps:
for the first resource in each candidate sequence, determining the corresponding passive exposure probability as a second preset value;
for a second resource in each candidate sequence, inputting the fifth feature vector into the third deep neural network layer to obtain the corresponding passive exposure probability.
7. The method of claim 4, wherein determining the estimated exposure probability and the estimated click rate of the target resource comprises:
for each candidate sequence, acquiring a sixth feature vector corresponding to each target resource based on a fourth embedding layer in a click rate prediction model, and acquiring a seventh feature vector corresponding to the first feature information based on a fifth embedding layer in the click rate prediction model; the sixth feature vector corresponding to each target resource is obtained based on the corresponding second feature information and the corresponding resource display position;
inputting each sixth feature vector into a self-attention layer connected with the fourth embedded layer in the click rate prediction model, so as to obtain a fusion feature vector corresponding to each target resource based on a self-attention mechanism;
inputting each fused feature vector and the seventh feature vector into a fourth deep neural network layer connected with the self-attention layer and the fifth embedded layer in the click rate prediction model to obtain the predicted click rate of each target resource in the candidate sequence.
8. The method according to claim 1, wherein the second characteristic information includes bid information of a corresponding target resource and a resource type to which the target resource belongs; determining a target sequence to be recommended from a plurality of candidate sequences based on the estimated exposure probability and the estimated click rate of each target resource in each candidate sequence, including:
for each candidate sequence, determining similarity between the target resources based on second feature information of the target resources;
determining a score corresponding to the candidate sequence by combining the estimated exposure probability, the estimated click rate, the offer information, the type of the resource and the similarity between the target resources in a preset mode;
determining the target sequence from the plurality of candidate sequences based on scores corresponding to the plurality of candidate sequences.
9. The method of claim 7, wherein the sequence generation model, the click through rate prediction model, and the exposure probability prediction model are trained by:
acquiring first sample characteristic information and a plurality of sample sequences which are respectively related to a plurality of sample users, wherein each sample sequence comprises N sample resources which are obtained by random sampling and respectively correspond to one resource display position, and acquiring second sample characteristic information of each sample resource; each sample resource is labeled according to the corresponding sample exposure probability and the sample sampling rate;
respectively training the initial click rate pre-estimation model and the initial exposure probability pre-estimation model based on the first sample characteristic information, the second sample characteristic information of the sample resources, the corresponding resource display positions, the corresponding sample exposure probability and the sample sampling rate to obtain the trained click rate pre-estimation model and the trained exposure probability pre-estimation model;
based on the plurality of first sample characteristic information and the plurality of second sample characteristic information, combining the trained click rate pre-estimation model and the trained exposure probability pre-estimation model, training the initial sequence generation model to obtain the trained sequence generation model.
10. An apparatus for recommending resources, said apparatus comprising:
the resource recommendation system comprises a first acquisition module, a second acquisition module and a resource recommendation module, wherein the first acquisition module is used for acquiring a resource recommendation request, the request comprises first characteristic information related to a user and the number N of resource display positions, and N is an integer greater than 1;
the generating module is used for acquiring second characteristic information of a plurality of candidate resources and generating a plurality of candidate sequences based on the first characteristic information and the second characteristic information; each candidate sequence comprises N target resources which are selected from a plurality of candidate resources and respectively correspond to one resource display position;
the determining module is used for determining the estimated exposure probability and the estimated click rate of the target resource for at least one target resource in each candidate sequence based on the corresponding second characteristic information and the corresponding resource display position, and the second characteristic information and the corresponding resource display position of at least one other target resource except the target resource in the candidate sequence and the first characteristic information;
and the recommending module is used for determining a target sequence to be recommended from the candidate sequences based on the estimated exposure probability and the estimated click rate of each target resource in each candidate sequence, and recommending resources according to the resource display position corresponding to each target resource in the target sequence.
11. The apparatus of claim 10, wherein the generating module comprises:
a first obtaining unit, configured to obtain, for each candidate resource, a confidence level of display of the candidate resource at each resource display position based on the first feature information and the plurality of second feature information through a sequence generation model;
and the generating unit is used for selecting a plurality of target resources corresponding to each resource display position from the plurality of candidate resources based on the confidence degree of each candidate resource displayed at each resource display position through a preset selection mode, and generating a plurality of candidate sequences based on the plurality of target resources corresponding to the N resource display positions.
12. The apparatus of claim 11, wherein the sequence generative model comprises a first embedding layer, a first deep neural network layer, and a normalization layer connected in sequence; the first obtaining unit is configured to:
acquiring a first feature vector corresponding to the first feature information and a second feature vector corresponding to each piece of second feature information based on the first embedding layer, and acquiring global feature vectors corresponding to all the candidate resources based on each second feature vector;
for each candidate resource, splicing the first feature vector, the global feature vector and the corresponding second feature vector to obtain a corresponding first spliced vector;
inputting the first stitching vector corresponding to each candidate resource into the first deep neural network layer to obtain a first initial confidence of each candidate resource displayed at each resource display position;
and inputting each first initial confidence degree into the normalization layer to obtain each confidence degree corresponding to each candidate resource.
13. The apparatus according to claim 11 or 12, wherein the estimated exposure probability comprises an active exposure probability associated with an active action of a user, wherein N of the resource display positions are arranged in sequence; the determining module includes:
a first determining unit, configured to determine, for a first resource in each candidate sequence, the corresponding active exposure probability as a first preset value; the first resource is a target resource corresponding to a resource display position arranged at the head in the candidate sequence;
a second determining unit, configured to determine, for a second resource in each candidate sequence, the active exposure probability of the second resource based on corresponding second feature information and a corresponding resource display position, second feature information and a corresponding resource display position of a third resource in the candidate sequence, and the first feature information through an exposure probability pre-estimation model; the second resource is other target resources except the first resource in the candidate sequence; the third resource is a target resource of which the corresponding resource display position in the candidate sequence is arranged before the second resource.
14. The apparatus of claim 13, wherein the pre-exposure probability prediction model comprises a second embedding layer, a third embedding layer, a splicing layer connected to the second embedding layer and the third embedding layer, and a second deep neural network layer connected to the splicing layer;
the second determining unit is configured to:
acquiring a third feature vector and a fourth feature vector based on the second embedded layer, and acquiring a fifth feature vector corresponding to the first feature information based on the third embedded layer; the third feature vector is obtained based on second feature information of the second resource and a corresponding resource display position; the fourth feature vector is obtained based on second feature information of the third resource and a corresponding resource display position;
inputting the third feature vector, the fourth feature vector and the fifth feature vector into the splicing layer to obtain a corresponding second splicing vector;
inputting the second stitching vector into the second deep neural network layer to obtain the active exposure probability of the second resource.
15. The apparatus of claim 14, wherein the estimated exposure probability further comprises a passive exposure probability independent of the user's active behavior; the exposure probability pre-estimation model further comprises a third deep neural network layer connected with the third embedded layer;
the determining module further comprises:
a third determining unit, configured to determine, for the first resource in each candidate sequence, the corresponding passive exposure probability as a second preset value;
a second obtaining unit, configured to, for a second resource in each candidate sequence, input the fifth feature vector into the third deep neural network layer to obtain the corresponding passive exposure probability.
16. The apparatus of claim 13, wherein the determining module comprises:
a third obtaining unit, configured to obtain, for each candidate sequence, a sixth feature vector corresponding to each target resource based on a fourth embedding layer in a click rate prediction model, and obtain a seventh feature vector corresponding to the first feature information based on a fifth embedding layer in the click rate prediction model; the sixth feature vector corresponding to each target resource is obtained based on the corresponding second feature information and the corresponding resource display position;
a fourth obtaining unit, configured to input each sixth feature vector into an attention layer connected to the fourth embedded layer in the click rate prediction model, so as to obtain a fusion feature vector corresponding to each target resource based on an attention mechanism;
a fifth obtaining unit, configured to input each of the fused feature vectors and the seventh feature vector into a fourth deep neural network layer connected to the self-attention layer and the fifth embedded layer in the click rate prediction model, so as to obtain the predicted click rate of each of the target resources in the candidate sequence.
17. The apparatus according to claim 10, wherein the second characteristic information includes a bid information of a corresponding target resource and a resource type to which the target resource belongs; the recommendation module comprises:
a fourth determining unit, configured to determine, for each candidate sequence, a similarity between the target resources based on second feature information of the target resources;
a fifth determining unit, configured to determine, in a preset manner, a score corresponding to the candidate sequence by combining the estimated exposure probability of each target resource, the estimated click rate, the bid information, the resource type to which the target resource belongs, and a similarity between the target resources;
a sixth determining unit, configured to determine the target sequence from the multiple candidate sequences based on scores corresponding to the multiple candidate sequences.
18. The apparatus of claim 16, further comprising:
a second obtaining module, configured to obtain first sample feature information and a plurality of sample sequences that are respectively related to a plurality of sample users, where each sample sequence includes N sample resources that are obtained by random sampling and each of which corresponds to one of the resource display positions, and obtain second sample feature information of each of the sample resources; each sample resource is labeled according to the corresponding sample exposure probability and the sample sampling rate;
the first training module is used for training the initial click rate estimation model and the initial exposure probability estimation model respectively based on the plurality of first sample characteristic information, the second sample characteristic information of the plurality of sample resources, the corresponding resource display positions, the corresponding sample exposure probabilities and the sample sampling rates so as to obtain the trained click rate estimation model and the trained exposure probability estimation model;
and the second training module is used for training the initial sequence generation model based on the plurality of first sample characteristic information and the plurality of second sample characteristic information and by combining the trained click rate estimation model and the trained exposure probability estimation model to obtain the trained sequence generation model.
19. An electronic device, comprising:
at least one processor; and
a memory communicatively coupled to the at least one processor; wherein the content of the first and second substances,
the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-9.
20. A non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform the method of any one of claims 1-9.
CN202211415202.3A 2022-11-11 2022-11-11 Resource recommendation method and device, electronic equipment and storage medium Pending CN115905690A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211415202.3A CN115905690A (en) 2022-11-11 2022-11-11 Resource recommendation method and device, electronic equipment and storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211415202.3A CN115905690A (en) 2022-11-11 2022-11-11 Resource recommendation method and device, electronic equipment and storage medium

Publications (1)

Publication Number Publication Date
CN115905690A true CN115905690A (en) 2023-04-04

Family

ID=86490829

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211415202.3A Pending CN115905690A (en) 2022-11-11 2022-11-11 Resource recommendation method and device, electronic equipment and storage medium

Country Status (1)

Country Link
CN (1) CN115905690A (en)

Similar Documents

Publication Publication Date Title
US10958748B2 (en) Resource push method and apparatus
US20220351016A1 (en) Presentation module for webinterface production and deployment system
CN107609152B (en) Method and apparatus for expanding query expressions
EP3862893A1 (en) Recommendation model training method, recommendation method, device, and computer-readable medium
US11403532B2 (en) Method and system for finding a solution to a provided problem by selecting a winner in evolutionary optimization of a genetic algorithm
US8543518B2 (en) Deducing shadow user profiles for ad campaigns
EP4181026A1 (en) Recommendation model training method and apparatus, recommendation method and apparatus, and computer-readable medium
US20150052003A1 (en) Providing Personalized Item Recommendations Using Scalable Matrix Factorization With Randomness
CN110210882A (en) Promote position matching process and device, promotion message methods of exhibiting and device
CN109783539A (en) Usage mining and its model building method, device and computer equipment
CN113837842A (en) Commodity recommendation method and equipment based on user behavior data
KR102295799B1 (en) Metaverse service management server for managing the metaverse service to change interface design of metaverse environment based on payment record, and the operating method thereof
CN111460384A (en) Policy evaluation method, device and equipment
CN113761348A (en) Information recommendation method and device, electronic equipment and storage medium
US11755979B2 (en) Method and system for finding a solution to a provided problem using family tree based priors in Bayesian calculations in evolution based optimization
CN102760127B (en) Method, device and the equipment of resource type are determined based on expanded text information
CN116823410B (en) Data processing method, object processing method, recommending method and computing device
Lettner et al. Mobile interaction analysis: towards a novel concept for interaction sequence mining
JP5624689B2 (en) Determining the degree of agreement between information categories and display information
WO2021021433A1 (en) Machine learning system to predict causal treatment effects of actions performed on websites or applications
CN112989174A (en) Information recommendation method and device, medium and equipment
CN111507471B (en) Model training method, device, equipment and storage medium
CN115905690A (en) Resource recommendation method and device, electronic equipment and storage medium
WO2020211616A1 (en) Method and device for processing user interaction information
CN112567415B (en) Modeling distribution of digital components through a network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination