CN112785391A

CN112785391A - Recommendation processing method and device, intelligent device and storage medium

Info

Publication number: CN112785391A
Application number: CN202110144785.XA
Authority: CN
Inventors: 张懿; 甘露; 吴伟佳; 洪庚伟; 李羽
Original assignee: Weimin Insurance Agency Co Ltd
Current assignee: Weimin Insurance Agency Co Ltd
Priority date: 2021-02-02
Filing date: 2021-02-02
Publication date: 2021-05-11
Anticipated expiration: 2041-02-02
Also published as: CN112785391B

Abstract

The embodiment of the application discloses a recommendation processing method and device, intelligent equipment and a storage medium. The method is applied to intelligent equipment, a ranking model obtained by pre-training is configured on the intelligent equipment, and the ranking model comprises a first network layer, a second network layer and a third network layer. The method comprises the following steps: and acquiring fusion characteristic data of each element to be recommended in the element set to be recommended. And processing the fused feature data through a first network layer to generate a first feature vector and a second feature vector corresponding to each element to be recommended. And respectively processing the first eigenvector and the second eigenvector corresponding to each element to be recommended through a first subnetwork of a target type included in a first network in the second network layer and a second network in the second network layer to obtain a first output vector and a second output vector. And processing the first output vector and the second output vector through a third network layer to obtain the predicted value of each element to be recommended. By adopting the embodiment of the application, the recommendation precision can be improved.

Description

Recommendation processing method and device, intelligent device and storage medium

Technical Field

The present application relates to the field of artificial intelligence technologies, and in particular, to a recommendation processing method and apparatus, an intelligent device, and a storage medium.

Background

With the rapid development of the internet, the application scenarios of the item recommendation method are more and more. The related recommendation provided by the item recommendation method can help the user to better select satisfactory items, and meanwhile, the income improvement of an e-commerce platform can be promoted.

At present, for a recommendation method in a multi-page recommendation scene, related technologies mostly recommend to a user according to elements such as commodities, services and the like recently clicked or browsed by the user, for example, the user often pays attention to office stationery recently, when the user needs to recommend corresponding elements to be recommended, the office stationery often ranks forward on a page, and the user can find commodity elements such as the office stationery more quickly. How to better rank various elements such as commodities and services becomes a hot issue of research when recommending the elements.

Disclosure of Invention

The embodiment of the application provides a recommendation processing method and device, an intelligent device and a storage medium, and can better realize the sequencing of elements to be recommended.

The embodiment of the application provides a recommendation processing method, which is applied to intelligent equipment, wherein a ranking model obtained by pre-training is configured on the intelligent equipment, the ranking model comprises a first network layer, a second network layer and a third network layer, and the method comprises the following steps:

acquiring fusion characteristic data of each element to be recommended in the element set to be recommended;

processing the fusion feature data corresponding to each element to be recommended through a first network layer to generate a first feature vector and a second feature vector corresponding to each element to be recommended;

processing a first feature vector corresponding to each element to be recommended through a first sub-network of a target type included in a first network in a second network layer to obtain a first output vector, wherein the first network includes a plurality of types of first sub-networks, and the target type is matched with the type to which the set of elements to be recommended belongs;

processing a second feature vector corresponding to each element to be recommended through a second network in a second network layer to obtain a second output vector;

and processing the first output vector and the second output vector through a third network layer to obtain a predicted value of each element to be recommended in the element set to be recommended, and determining the recommended arrangement sequence of each element to be recommended in the element set to be recommended according to the predicted values.

The embodiment of the application provides a recommendation processing device, the device is provided with a ranking model obtained by pre-training, the ranking model comprises a first network layer, a second network layer and a third network layer, and the device comprises:

the fusion characteristic data acquisition module is used for acquiring fusion characteristic data of each element to be recommended in the element set to be recommended;

the fusion feature data processing module is used for processing fusion feature data corresponding to each element to be recommended through a first network layer to generate a first feature vector and a second feature vector corresponding to each element to be recommended;

a first output vector determining module, configured to process, through a first sub-network of a target type included in a first network in a second network layer, a first feature vector corresponding to each element to be recommended to obtain a first output vector, where the first network includes a plurality of types of first sub-networks, and the target type is matched with a type to which the set of elements to be recommended belongs;

the second output vector determining module is used for processing a second feature vector corresponding to each element to be recommended through a second network in a second network layer to obtain a second output vector;

and the predicted value determining module is used for processing the first output vector and the second output vector through a third network layer to obtain a predicted value of each element to be recommended in the element set to be recommended, and determining the recommended arrangement sequence of each element to be recommended in the element set to be recommended according to the predicted values.

The embodiment of the application provides a smart device which comprises a processor, a memory and a transceiver, wherein the processor, the memory and the transceiver are connected with each other. The memory is used for storing a computer program enabling the smart device to perform the method provided by the first aspect, the computer program comprising program instructions, the processor and the transceiver being configured to invoke the program instructions to perform the method provided by the first aspect.

Embodiments of the present application provide a computer-readable storage medium storing a computer program, the computer program comprising program instructions, which, when executed by a processor, cause the processor to perform the method provided above.

In the embodiment of the application, two sets of networks are designed to form a recommendation model, for the obtained fusion feature data of the elements to be recommended, a first feature vector and a second feature vector corresponding to each element to be recommended are generated through a first network layer, the first feature vector corresponding to each element to be recommended is processed through a first sub-network which is included in a first network in a second network layer and is related to the type of each element to be recommended to obtain a first output vector, the second feature vector corresponding to each element to be recommended is processed through a second network in the second network layer to obtain a second output vector, the first output vector and the second output vector are processed through a third network layer to obtain a predicted value of each element to be recommended in a set of elements to be recommended, different first sub-networks are configured for different sets of elements to be recommended in the first network, therefore, the elements to be recommended are subjected to targeted analysis processing based on types, the elements to be recommended are subjected to global processing in the second network, and finally sorting is realized through the third network layer, so that the elements to be recommended can be sorted more accurately, the recommendation efficiency is improved, and the automatic and intelligent requirements of users on commodity and service recommendation are met.

Drawings

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

FIG. 1 is a system architecture diagram according to an embodiment of the present application;

FIG. 2 is a schematic structural diagram of a ranking model provided in an embodiment of the present application;

fig. 3 is a schematic flowchart of a recommendation processing method provided in an embodiment of the present application;

FIG. 4 is a diagram illustrating verification results of ranking models provided by an embodiment of the present application;

FIG. 5 is a schematic diagram of the verification result of the AB test provided in the embodiments of the present application;

fig. 6 is a schematic structural diagram of a recommendation processing apparatus according to an embodiment of the present application;

fig. 7 is another schematic structural diagram of a recommendation processing apparatus according to an embodiment of the present application;

fig. 8 is a schematic structural diagram of an intelligent device provided in an embodiment of the present application.

Detailed Description

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

The recommendation processing method provided by the embodiment of the application can be applied to intelligent equipment, a ranking model obtained by pre-training is configured on the intelligent equipment, and the ranking model comprises a first network layer, a second network layer and a third network layer. The smart device includes, but is not limited to, a server, a terminal device, and the like, where the terminal device includes, but is not limited to, a smart phone, a tablet computer, a notebook computer, a desktop computer, and the like, and the present disclosure is not limited thereto. For convenience of description, the following description will be given taking an intelligent device as an example. It can be understood that the intelligent device configured with the ranking model can be used for recommending each element to be recommended included in any page to be recommended in various online recommendation systems. For example, the online recommendation system may be a commodity recommendation system, a service recommendation system, a song recommendation system, a video recommendation system, and the like, and is determined according to an actual application scenario, which is not limited herein. The service recommendation system may be a system for recommending insurance services, and the like, which is not limited herein. For convenience of description, the object to be recommended in the recommendation system is referred to as an element to be recommended, or an item to be recommended (item), and the like. The item to be recommended is a broad-sense item, and is a general term of all recommendable objects.

Referring to fig. 1, fig. 1 is a schematic diagram of a system architecture according to an embodiment of the present disclosure. As shown in fig. 1, the server 100a, the server 100b, and the terminal apparatus 101 may be connected to each other via a network. The terminal device shown in fig. 1 may include a mobile phone, a tablet computer, a notebook computer, a palm computer, a Mobile Internet Device (MID), a wearable device (e.g., a smart watch, a smart bracelet, etc.), and the like, which are not limited herein. It should be understood that the server 100a may perform data transmission with the server 100b, and the server 100a and the server 100b may also perform data transmission with the terminal device 101, respectively. For convenience of description, the application programs with various functionalities loaded on the terminal device may be exemplified by the first application. Optionally, the first application may be a video application, a music playing application, a game application, a shopping application, or the like, or the first application may also be an applet, a web page, or the like, which is not limited herein.

It is understood that the server 100a and the server 100b may be a local server of the first application, a remote server (e.g., a cloud server), and the like, and are not limited herein. Among other things, the server 100a may be used to store various types of data (e.g., user characteristic data of a registered user of the first application, element characteristic data of each element to be recommended, etc.) and information related to the first application. The server 100b may be configured to obtain various feature data required by a certain page recommendation scenario from the server 100a, and perform processing based on the various feature data to generate a recommendation list for sending to the terminal device for displaying. That is, when the first application in the terminal device sends a data acquisition request to the server 100b, the server 100b may return a data condition (e.g., a recommendation list as in fig. 1) to the first application in the terminal device based on the received data acquisition request to update the interface display condition of the first application in the terminal device.

The method in the embodiment of the application can be applied to intelligent equipment, wherein a ranking model obtained by pre-training is configured on the intelligent equipment, and the ranking model comprises a first network layer, a second network layer and a third network layer. Specifically, fusion feature data of each element to be recommended in the element set to be recommended is obtained. And processing the fusion feature data corresponding to each element to be recommended through the first network layer to generate a first feature vector and a second feature vector corresponding to each element to be recommended. And processing the first feature vector corresponding to each element to be recommended through a first sub-network of a target type included in a first network in a second network layer to obtain a first output vector, wherein the first network includes a plurality of types of first sub-networks, and the target type is matched with the type to which the element set to be recommended belongs. And processing the second eigenvector corresponding to each element to be recommended through a second network in a second network layer to obtain a second output vector. And processing the first output vector and the second output vector through a third network layer to obtain a predicted value of each element to be recommended in the element set to be recommended, and determining the recommended arrangement sequence of each element to be recommended in the element set to be recommended according to the predicted values. By adopting the embodiment of the application, the recommendation precision can be improved.

It can be understood that please refer to fig. 2, fig. 2 is a schematic structural diagram of the ranking model provided in the embodiment of the present application. As shown in fig. 2, the ranking model in the embodiment of the present application may include a first network layer, a second network layer, and a third network layer. The first network layer comprises an embedded layer network and a splicing collocation layer, and the second network layer comprises a routing network, a first network TSNA and a second network GEN. Wherein the first network TSNA comprises m first sub-networks, such as the first sub-network TSN in fig. 2₀…, first sub-network TSN_m-1. The second network GEN comprises n second sub-networks, such as the second sub-network g in fig. 2₀…, second subnetwork g_n-1. Wherein m and n are integers greater than 1. The value of m is the same as the number of the page types of the recommended pages in the multi-page recommendation scene. That is, one first sub-network in the first network is used to learn the characteristics of the elements to be recommended in one type of recommendation page. The third network layer includes attention unit queues, a concatenation collocation layer, a routing network, and a third network (the third network may also be called a fine tuning network or a fine tuning layer), which includes m third sub-networks, such as the third sub-network MLP in fig. 2₁…, third sub-network MLP_m-1。

The first network layer is mainly used for processing fusion feature data of each element to be recommended included in the target recommendation page to obtain a first feature vector and a second feature vector. The first network TSNA in the second network layer is mainly used for processing the first eigenvector corresponding to each element to be recommended to obtain a first output vector, and the second network GEN in the second network layer is mainly used for processing the second eigenvector of each element to be recommended to obtain a second output vector corresponding to each element to be recommended. And the third network layer is mainly used for processing the first output vector and the second output vector to obtain a predicted value of the element to be recommended.

It is understood that a first sub-network comprised by the first network in the second network layer in the present application may be used to independently learn the characteristics of the elements to be recommended comprised in one type of recommendation page. Therefore, the number of the first sub-networks in the first network is the same as the number of the page types of the recommendation pages included in the multi-page recommendation scenario, that is, each type of recommendation page has a unique first sub-network, which can be used to learn the feature distribution of the elements to be recommended included in the corresponding recommendation page. That is, assuming that the multi-page recommendation scene includes m types of recommendation pages, in the training phase of the ranking model, for one first sub-network in the first network, the one first sub-network may be trained from one sample data corresponding to one type of recommendation page. Therefore, m types of sample data corresponding to the m types of recommendation pages can be trained to obtain m first sub-networks. The sample data corresponding to one type of recommended page comprises page feature data of the one type of recommended page, user feature data of a sample user and element feature data of a sample element to be recommended included in the one type of recommended page.

Correspondingly, in order to learn the correlation characteristics among the recommended pages in the multi-page recommendation scene, so as to improve the learning effect and further improve the recommendation precision, the second network of the second network layer in the application can be obtained by training according to sample data corresponding to the recommended pages comprising multiple types in the multi-page recommendation scene. For example, if a multi-page recommendation scene includes m types of recommendation pages, in the training phase of the ranking model, the second network may be obtained by training according to m types of training data corresponding to the m types of recommendation pages. The training data comprises page feature data of one type of recommended page, user feature data of a sample user and element feature data of a sample element to be recommended included in the one type of recommended page.

The specific implementation process of each network layer will be described in detail in each step in the following embodiments, and is not described in detail here.

The method and the apparatus provided by the embodiments of the present application will be described in detail with reference to fig. 3 to 8, respectively.

Referring to fig. 3, fig. 3 is a schematic flow chart of a recommendation processing method according to an embodiment of the present application. The method provided by the embodiment of the application can comprise the following steps S301 to S305:

s301, obtaining fusion characteristic data of each element to be recommended in the element set to be recommended.

In some possible embodiments, the methods provided in the present application may be applied to smart devices, wherein the smart devices include, but are not limited to, servers, terminal devices, and the like, wherein the terminal devices include, but are not limited to, smart phones, tablet computers, notebook computers, desktop computers, and the like, and are not limited thereto. For convenience of description, the following description will take an intelligent device as an example. Specifically, the intelligent device may obtain fusion feature data of each element to be recommended in the set of elements to be recommended, where the fusion feature data corresponding to one element to be recommended may be generated according to user feature data of a target recommendation user, the element feature data of the element to be recommended, and page feature data of a target recommendation page. That is to say, to generate the fused feature data of each element to be recommended in the element set to be recommended, the associated features of the element set to be recommended may be obtained first. Wherein, the association characteristics of the element set to be recommended include: any one or more of user characteristic data of a target recommending user, page characteristic data of a target recommending page corresponding to the element set to be recommended, and element characteristic data of each element to be recommended in at least two elements to be recommended included in the target recommending page. Furthermore, the associated features of the element set to be recommended are processed based on the first network layer in the ranking model, and fusion feature data corresponding to each element to be recommended can be generated.

For example, for any element to be recommended in the element set to be recommended, the fusion feature data corresponding to the element to be recommended can be generated by splicing the element feature data of the element to be recommended, the user feature data of the target recommendation user, and the page feature data of the target recommendation page. The splicing mode of the user characteristic data, the page characteristic data and the element characteristic data can be determined according to the actual application scene, and is not limited herein. For example, if the target recommendation user is a user1, the user feature data corresponding to the user1 is feature1, the target recommendation page is a page scenario1, the page feature data corresponding to the page scenario1 is feature2, and the element feature data of the element to be recommended a in the page 1 is feature3, then the fusion feature data Ra corresponding to the element to be recommended a [ feature1, feature2, feature3], or the fusion feature data Ra corresponding to the element to be recommended a [ feature1, feature3, feature2], or the fusion feature data Ra corresponding to the element to be recommended a [ feature2, feature1, feature3], or the fusion feature data Ra corresponding to the element to be recommended a [ feature2, feature3, feature3, feature data, or the fusion feature data Ra corresponding to the element to be recommended a [ feature3, feature 465, or the fusion feature data Ra corresponding to the element to be recommended a [ feature Ra ],4648, feature2, feature1], and the like, without limitation.

The target recommending user is a directional object for the element recommendation/article recommendation. For example, if the item recommendation is directed to the user1, the user1 is the target recommendation user, and if the item recommendation is directed to the user2, the user2 is the target recommendation user. It is understood that the target recommending user is usually a registered user in the recommending system, wherein the registered user may be a newly registered user, i.e. a user who has not performed purchasing behavior or made an order record, or may also be an old registered user, i.e. a user who has performed purchasing behavior or made an order record, etc., without limitation. Wherein the user characteristic data of the target recommendation user comprises at least one of user gender, user age, user browsing history, user purchasing history and the like.

And the recommendation page corresponding to the element set to be recommended is a target recommendation page. That is to say, at least two elements to be recommended in the element set to be recommended in the application are elements in the target recommendation page recommendation scene. For example, assuming that the multi-page recommendation scenario includes a special sell page, a payment completion page, a policy list page, and a child product page, the target recommendation page may be the special sell page, or the payment completion page, or the policy list page, or the child product page. The page feature data of the target recommendation page may include at least one of a page identifier, a user browsing duration, a user location, and the like. The element to be recommended can also be called an article to be recommended, and the element characteristic data of the element to be recommended can be called article characteristic data of the article to be recommended. It should be understood that the item characteristic data of the item to be recommended may include at least one of an item identification, an item type, an item price, an item manufacturer, and the like.

For example, assume that the target recommendation page is a policy list page, wherein the elements to be recommended included in the policy list page are the respective risk types, and wherein the risk type feature data of each risk type may include at least one of a name of the risk type, a covered disease type, a price, an underwriting company, a premium, a minimum underwriting age, a maximum underwriting age, and the like. The user characteristic data of the target recommendation user can comprise the gender of the user, the age of the user, the annual insurance premium of the user, the payment of certain insurance policies, the failure insurance policies, the time for participating in health activities, whether the user is a public number attention user, whether the user is a vehicle user, whether the user is a new user, the click behavior in the last 30 days and the like. The page characteristic data of the policy list page may include page identification, user location, etc.

S302, processing the fusion feature data corresponding to each element to be recommended through the first network layer, and generating a first feature vector and a second feature vector corresponding to each element to be recommended.

In some possible embodiments, the fused feature data corresponding to each element to be recommended is processed by the first network layer, so that a first feature vector and a second feature vector corresponding to each element to be recommended can be generated. It is understood that the fused feature data in the present application may be feature data in LIBSVM format. That is, target recommendationsFor any element to be recommended in the page, the fused feature data corresponding to the element to be recommended may be a multi-heat vector. For example, gender "male" may be translated into such a form (gender _ male: 1) for gender in the user characteristic data, and for example, for browsing history in the user characteristic data, the browsing history feature may be constructed as (browsing history _ item a:1, browsing history _ item b:1, browsing history _ item c:0) assuming that the target recommends that the product browsed by the user is item a, item b, and the product not browsed is item c. Therefore, the fused feature data corresponding to the element to be recommended can be expressed as

Wherein the superscript j represents a recommendation page, and the subscript i represents the ith element to be recommended included in the recommendation page j. Therefore, fusion feature data corresponding to each element to be recommended included in the recommendation page

Feature data can be fused based on a vector dictionary (i.e., the first preset vector of the present application)

Conversion to dense vectors

In particular, the constructed dense vector may be represented as:

where emb represents a vector dictionary and x represents a dot product (i.e., multiplication of elements in a vector). Wherein dense vectors are constructed

Then, for the input of the first sub-network of the target type included in the first network in the second network layer and the input of the second network, the first network layer will respectively correspond to the dense vectors

And performing different processing to generate a first feature vector and a second feature vector, and further taking the first feature vector as an input of a first sub-network of a target type included by a first network in a second network layer, and taking the second feature vector as an input of a second network in the second network layer. Specifically, for the input of the first sub-network of the target type included in the first network in the second network layer, the first network layer can be processed by the dense vector pair

Summing by rows to obtain a first feature vector, which may be noted as a first feature vector for convenience of description

For the input of the second network in the second network layer, the first network layer can be used for processing the dense vector

The second eigenvector can be recorded as a second eigenvector for convenient description

For example, assume a dense vector

Has a dimension of [ N, M]The first network then comprises the input variables (i.e. the first eigenvector) of the first subnetwork of the target type

) Dimension of [1, M]Input variables of the second network (i.e. second eigenvectors)

) Dimension of [1, N M]Wherein N and M are integers greater than 1.

For example, assume the fusion feature in this applicationData is

The first preset vector is obtained

Dense vector based conversion formula

Corresponding dense vectors can be obtained as

Thus, by aligning dense vectors

Summing by rows to obtain a first eigenvector

By aligning dense vectors

The second eigenvector is obtained by opening according to the row and then splicing

S303, processing the first feature vector corresponding to each element to be recommended through a first sub-network of the target type included in the first network in the second network layer to obtain a first output vector.

In some possible embodiments, the first output vector may be obtained by processing the first feature vector corresponding to each element to be recommended through a first subnetwork of a target type included in the first network in the second network layer. The first network comprises a plurality of types of first sub-networks, and the target type is matched with the type to which the element set to be recommended belongs. That is, the plurality of types of the first sub-networks referred to in the embodiments of the present application may be understood as a plurality of page types corresponding to the first sub-networks, and one page type corresponds to one first sub-network. For example, assuming that a multi-page recommendation scenario includes 4 recommendation pages, the number of first subnetworks in the first network is 4 (i.e., m is 4). The target type may be a type determined according to a page identifier included in the page feature data of the target recommendation page.

It is to be understood that the first sub-network in the embodiment of the present application may be a multi-layer perceptron (MLP), or may also be Wide & Deep (a network structure), or may also be Deep fm (a network structure), and the like, which is determined according to an actual application scenario, and is not limited herein. For convenience of description, the first sub-network is taken as an MLP in the embodiments of the present application. That is, the first network in the second network layer is a set of MLPs for the page-oriented recommendation scenario. One MLP in the first network is used for learning the feature representation of each element to be recommended included in one page recommendation scene. In general, in the embodiment of the present application, the number of MLPs included in the first network is the same as the number of page types of recommended pages included in a multi-page recommendation scene, that is, each recommended page has an independent MLP. For example, if a recommendation system includes 4 recommendation pages, which are a special sale page, a payment completion page, a policy list page, and a child product page, the 4 recommendation pages correspond to 4 MLPs, and one MLP is used to independently learn the feature representation of the element to be recommended included in the corresponding recommendation page, so as to capture the corresponding data distribution.

In order to realize the purpose of learning the data distribution of each page recommendation scene respectively, the routing network for selective activation is designed in the embodiment of the application, the routing network can determine the target type according to the page feature data of the target recommendation page, and then the first feature vector corresponding to each element to be recommended in the target recommendation page is guided to the MLP corresponding to the target type for calculation. Specifically, the routing network may guide the first feature vector corresponding to each element to be recommended in the target recommendation page to the MLP corresponding to the target recommendation page for calculation according to the page identifier included in the page feature data of the target recommendation page, so as to obtain the first output vector. Wherein the output of the first network (i.e. the first output vector in the embodiment of the present application) can be calculated by the following formula:

wherein the content of the first and second substances,

a first output vector is represented that is a first output vector,

represents the output of the h-th MLP.

S304, processing the second feature vector corresponding to each element to be recommended through a second network in a second network layer to obtain a second output vector.

In some possible embodiments, the second feature vector corresponding to each element to be recommended is processed by a second network in the second network layer, so as to obtain a second output vector. It is understood that the second network in the second network layer may comprise n second subnetworks, n being an integer greater than 1. Wherein, the processing the second eigenvector corresponding to each element to be recommended through the second network in the second network layer to obtain the second output vector may be: firstly, second eigenvectors corresponding to elements to be recommended are determined as n vector sets, then the n vector sets are respectively input into the n second sub-networks, and the corresponding vector sets are respectively processed through the second sub-networks to obtain second output vectors. Wherein, determining the second feature vector corresponding to the element to be recommended as n vector sets can be understood as: and splitting the second characteristic vector corresponding to the element to be recommended to obtain n vector sets. For example, assume that the dimension of the second feature vector is [1, 10 ]]And n is 3. Wherein, if the second feature vector is specifically:

then pass through pair

Splitting is carried out to obtain 3 vector sets respectively

It can be understood that the length of each vector set split in the embodiment of the present application may be determined according to an actual application scenario, and is not limited herein.

In some possible embodiments, the above processing the corresponding vector set by each second sub-network to obtain the second output vector may be: obtaining each output vector obtained by processing the corresponding vector set by each second sub-network respectively

Wherein k is a whole number and is greater than or equal to 0 and less than or equal to n-1. And further based on respective output vectors output by respective second sub-networks

A second output vector is generated.

Specifically, referring to fig. 2 together, the second network in the second network layer may be used to learn the feature representation globally, unlike the first network in the second network layer which is used to learn the feature representation of the element to be recommended independently based on the target type only at a time. In order to capture feature representations of different spaces, a feature extraction mode similar to a multi-head attention mechanism is designed for the second network. For the second network, the embodiment of the present application may design it as a queue composed of a set of second sub-networks, and therefore, the output of the second network (i.e. the second output vector in the present application) may be represented as:

wherein the content of the first and second substances,

a second output vector is represented that is,

and expressing the output of the kth second sub-network, wherein the value range of k is more than or equal to 0 and less than or equal to n-1, and k is an integer. n denotes the number of second sub-networks, i.e. the second output vector is composed of the output vectors of each of the n second sub-networks. Each of the second sub-networks included in the second network may also be an MLP or may be a Wide, similar to the first sub-network included in the first network&Deep (a network structure), or Deep fm (a network structure) may also be adopted, which is determined according to the actual application scenario, and is not limited herein. Understandably, in the second network, the second network comprises each second sub-network

Focusing only on the input vector (i.e. the second eigenvector in this application)

) And thus the features of different spaces can be more fully captured, a particular implementation can be represented by the following equation:

wherein, g₀Second feature vector of interest

In x₁～x_l0Partial input, g₁Second feature vector of interest

In x_l0+1～x_l1In partInput, analogize with others, g_n-1Second feature vector of interest

In

Partial input. The length of the eigenvalue in the second eigenvector concerned by different second subnetworks may be adjusted according to the actual application scenario, which is not limited herein.

S305, processing the first output vector and the second output vector through a third network layer to obtain a predicted value of each element to be recommended in the element set to be recommended.

In some possible embodiments, the third network layer processes the first output vector and the second output vector to obtain a predicted value of each element to be recommended in the set of elements to be recommended. Specifically, a weight vector may be obtained by processing the first output vector and the second output vector by the attention unit of the target type in the attention unit queue in the third network layer, wherein the weight vector includes n weight values. And processing the weight vector, the first output vector and the second output vector through a third network layer to obtain a predicted value of each element to be recommended in the element set to be recommended. The attention unit queue comprises a plurality of attention units of different types, and the target type is matched with the type of the element set to be recommended. The processing of the first output vector and the second output vector by the attention unit of the target type in the attention unit queue in the third network layer to obtain the weight vector may be: the first output vector and the second output vector are processed through the attention unit of the target type in the attention unit queue in the third network layer to obtain a process parameter vector, and then the weight vector is determined according to the process parameter vector, wherein the process parameter vector comprises n process parameters.

In some possible embodiments, the processing, by the third network layer, the weight vector, the first output vector, and the second output vector to obtain the predicted value of each element to be recommended in the set of elements to be recommended may be: and processing the weight vector, the first output vector and the second output vector through a splicing collocation network of a third network layer to obtain a predicted feature vector. Further, the predicted feature vector is processed through a third sub-network of the target type in a third network layer, and a predicted value corresponding to each element to be recommended can be determined. It is understood that the third network comprises a plurality of third sub-networks of different types, and the target type is matched with the type to which the element set to be recommended belongs.

Specifically, referring to fig. 2 together, the embodiment of the present application assists element recommendation of a current page recommendation scenario (i.e. a target recommendation page) by designing a set of attention unit queues in a third network layer for extracting useful recommendation knowledge. Similar to the first network, the embodiment of the present application may assign an attention unit to each page recommendation scenario for learning a specific extraction rule (i.e., a process parameter vector in the present application). That is, the application is implemented with respect to the first output vector by attention units of the target type in the attention unit queue in the second network layer

And a second output vector

Processing to obtain a process parameter vector a^jIn particular, a process parameter vector a^jSatisfies the following conditions:

wherein, a^jRepresenting a process parameter vector, process parameter vector a^jComprising n process parameters respectively

That is, the number of process parameters included in the process parameter vector is the same as the number of second subnets.

Representing a first output vector, W is an attention weight matrix of the attention unit,

output vectors representing data from respective second sub-networks

The output vector of the second network.

Further, embodiments of the present application may utilize a^jTo calculate a weight vector

Wherein the weight vector

Satisfies the following conditions:

wherein the content of the first and second substances,

is the weight vector that is calculated as a function of,

including n weight values, that is,

the number of weight values in (b) is the same as the number of second sub-networks comprised in the second network. Therefore, the first output vector is finally concatenated through the Concatenation collocation network in the third network layer

Second output vector

And a weight vector

And processing to obtain the prediction characteristic vector corresponding to the element to be recommended. Wherein the predicted feature vector satisfies:

wherein the content of the first and second substances,

a first output vector is represented that is a first output vector,

is a second output vector

Including the output vectors of the respective second sub-networks,

representing weight vectors

The respective weight values included, η, represent normalization parameters.

It is to be understood that in some possible embodiments, a third network is further included in the third network layer in the present application, wherein the third network includes m third subnetworks. It can be understood that, in order to implement the calculation of the logoss of each page recommendation scene, a routing network for selective activation is also designed in the third-layer network in the embodiment of the present application, and the routing network may determine a target type according to a page identifier in page feature data of a target recommendation page, and then guide a predicted feature vector corresponding to an element to be recommended in the target recommendation page to the third sub-network corresponding to the target type for calculation. It is to be understood that the third sub-network in the embodiment of the present application may be an MLP, or may also be Wide & Deep (a network structure), or may also be Deep fm (a network structure), and the like, which is determined according to an actual application scenario and is not limited herein. For convenience of description, the third sub-network is taken as an MLP in the embodiments of the present application. Wherein a third sub-network (i.e. an MLP) in the third network is used to calculate the logoss of a page recommendation scene to determine the final output result. As shown in fig. 2, the embodiment of the present application establishes an independent third sub-network for each page recommendation scenario, and without loss of generality, in the current design, the embodiment of the present application adopts MLP as a fine tuning network, and the final output result (i.e. the predicted value in the present application) can be represented by the following formula:

wherein the content of the first and second substances,

indicating the MLP corresponding to the page recommendation scenario j,

and representing a predicted value corresponding to the current page recommendation scene (namely, the target recommendation page). In MLP, the ReLU is selected as the activation unit in the embodiments of the present application, because it has obvious advantages in forward gradient transmission, reducing gradient disappearance and the like.

Finally, the embodiment of the present application calculates the final loginoss by using weighted cross entropy, and controls the complexity of the model by adding l1 regular, wherein the cross entropy function satisfies:

wherein the content of the first and second substances,

represents the cross entropy, β is a hyper-parameter, representing the weight of each page recommendation scenario, whose value is specified by the expert.

The presence of a real label is indicated,

denotes the predicted value, λ denotes the weight of the l1 regularization term, and Ω denotes the parameter l1 regularization of the model.

In the embodiment of the application, fusion characteristic data of each element to be recommended in the element set to be recommended is obtained. And processing the fusion feature data corresponding to each element to be recommended through the first network layer to generate a first feature vector and a second feature vector corresponding to each element to be recommended. And processing the first feature vector corresponding to each element to be recommended through a first sub-network of a target type included in a first network in a second network layer to obtain a first output vector, wherein the first network includes a plurality of types of first sub-networks, and the target type is matched with the type to which the element set to be recommended belongs. And processing the second eigenvector corresponding to each element to be recommended through a second network in a second network layer to obtain a second output vector. And processing the first output vector and the second output vector through a third network layer to obtain a predicted value of each element to be recommended in the element set to be recommended, and determining the recommended arrangement sequence of each element to be recommended in the element set to be recommended according to the predicted values. By adopting the embodiment of the application, the recommendation precision can be improved.

It can be understood that, in order to verify the effectiveness of the ranking model proposed in the present application, the embodiments of the present application verify the model effect by using actual data. Wherein. The data sets for verification are all derived from internet insurance sales data in real scenes, and comprise about 560 pieces of data, which mainly comprise data of the following four page recommendation scenes:

special sale page (scenario 1): for the recommended scene of a user with certain cognition, the sample number is 390W.

Payment completion page (scienio 2): the sample number is 18.3W for the recommended scenario of the user who has completed the product purchase.

Policy list page (scienio 3): the recommended scenes of the users have not been purchased yet, and the number of samples is 41W.

Daughter product page (scienio 4): sample number 115W for the product recommendation scenario prepared for children/children.

At the time of testing, the present example used 80% of the data in the data set for testing and 20% of the data in the data set for verification. Understandably, the AUC (a parameter for measuring the ranking ability of the model), the Recall @ N (Recall rate) and the NDCG (a parameter for measuring the relevance of the ranking) are selected as the evaluation indexes for measuring the performance of the model in the embodiment of the present application. The other 5 ranking models for comparison with the ranking model provided in the embodiment of the present application are respectively:

(1) logistic Regression (LR) model: during testing, the LR model is obtained mainly by adopting FTRL (a training mode) training.

(2) Depth and logistic regression (Wide & Deep) model: this structure was used as the middle feature interaction layer when tested.

(3) Bottom Shared Model (BSM) model: a modified multi-task recommendation network, wherein multiple tasks share the same embedding layer, and weighting processing is only carried out on the final fine tuning layer.

(4) Puncture network (cotet) model: the feature interaction layer adopts a feature interaction method of fixed weight sharing.

(5) Multi-gate of mix-of-experts (MMoE) model: the interaction layer adopts an expert network queue, and the input of the fine adjustment layer is determined by an input sample and a control gate.

In the test, the learning rate is 0.002, the weight β ═ 0.5, 0.1, 0.2, and the regular weight l1 is 0.1. Each network model is fine tuned to the optimal result. Referring to fig. 4, fig. 4 is a schematic diagram illustrating a verification result of each ranking model provided in the embodiment of the present application. The test results of the ranking model proposed in the present application (i.e., SAMN in fig. 4) and other 5 ranking models (e.g., LR, Wide & Deep, BSM, cent, MMoE in fig. 4) are shown in fig. 4. As can be seen from fig. 4, in the multi-page recommendation scenario, the recommendation model provided in the present application exceeds other 5 ranking models in the training set/test set on the three indexes (i.e., AUC, Recall @ N and NDCG), so that the effectiveness of the ranking model provided in the present application is demonstrated.

Meanwhile, the embodiment of the present application further performs an AB test for 7 days on the 4 page recommendation scenes, please refer to fig. 5, and fig. 5 is a schematic diagram of a verification result of the AB test provided in the embodiment of the present application. In the embodiment of the present application, online CR (conversion rate) is normalized, a baseline (tree model) is normalized to 1, and the verification result is shown in fig. 5: in the above 4 page recommendation scenarios, the present application compares with the baseline, and the average CR increase value in the 4 page recommendation scenarios is: 3.98%, 1.04%, 1.26% and 1.05%. The AB test results also demonstrate the effectiveness of the present application once again.

Referring to fig. 6, fig. 6 is a schematic structural diagram of a recommendation processing device according to an embodiment of the present application. The recommendation processing device is an intelligent device, a ranking model obtained by pre-training is configured on the intelligent device, and the ranking model comprises a first network layer, a second network layer and a third network layer. The device of the embodiment of the invention can be arranged in intelligent equipment such as an intelligent terminal, a server and the like, the intelligent terminal can be various intelligent mobile phones, tablet computers, intelligent wearable equipment, personal computers and the like, the server mainly refers to some commodity and service recommendation platforms, even servers provided by some social application platforms, and the recommendation processing device provided by the embodiment of the invention comprises:

the fusion characteristic data acquisition module 61 is used for acquiring fusion characteristic data of each element to be recommended in the element set to be recommended;

the fused feature data processing module 62 is configured to process the fused feature data corresponding to each element to be recommended through the first network layer, and generate a first feature vector and a second feature vector corresponding to each element to be recommended;

a first output vector determining module 63, configured to process, through a first sub-network of a target type included in a first network in a second network layer, a first feature vector corresponding to each element to be recommended to obtain a first output vector, where the first network includes a plurality of types of first sub-networks, and the target type is matched with a type to which the set of elements to be recommended belongs;

a second output vector determining module 64, configured to process, through a second network in a second network layer, a second feature vector corresponding to each element to be recommended to obtain a second output vector;

the predicted value determining module 65 is configured to process the first output vector and the second output vector through a third network layer to obtain a predicted value of each element to be recommended in the set of elements to be recommended, so as to determine a recommended arrangement sequence of each element to be recommended in the set of elements to be recommended according to the predicted values.

Referring to fig. 7, fig. 7 is another schematic structural diagram of a recommendation processing device according to an embodiment of the present application. The device provided by the embodiment of the invention can be arranged in intelligent equipment such as an intelligent terminal, a server and the like, the intelligent terminal can be various intelligent mobile phones, tablet computers, intelligent wearable equipment, personal computers and the like, the server mainly refers to some commodities, service recommendation platforms and even servers provided by some social application platforms, and the device comprises the following structure.

In some possible embodiments, the fused feature data obtaining module 61 includes:

an associated feature obtaining unit 611, configured to obtain an associated feature of the to-be-recommended element set, where the associated feature includes: any one or more of user characteristic data of a target recommending user, page characteristic data of a target recommending page corresponding to an element set to be recommended, and article characteristic data of each element to be recommended in at least two elements to be recommended included in the target recommending page;

a fusion feature data generating unit 612, configured to generate, by the first network layer, fusion feature data corresponding to each element to be recommended according to the associated features of the set of elements to be recommended.

In some possible embodiments, the dimension of the fused feature data is [1 × N ]; the fused feature data processing module 62 includes:

a dense vector determining unit 621, configured to obtain a first preset vector, and generate a dense vector according to the fused feature data and the first preset vector, where a dimension of the first preset vector is [ M, N ], and a dimension of the dense vector is [ N, M ];

and a feature vector generation unit 622, configured to determine, according to the dense vector, a first feature vector and a second feature vector corresponding to each element to be recommended, where a dimension of the first feature vector is [1, M ], and a dimension of the second feature vector is [1, N × M ].

In some possible embodiments, the second network in the second network layer includes n second subnetworks, n being an integer greater than 1; the second output vector determination module 64 includes:

a vector set determining unit 641 configured to determine the second feature vector as n vector sets;

a second output vector determining unit 642, configured to input the n vector sets into the n second sub-networks, respectively, so as to process the corresponding vector sets through each second sub-network, respectively, to obtain second output vectors.

In some possible embodiments, the predicted value determination module 65 includes:

a weight vector determining subunit 651, configured to process the first output vector and the second output vector through an attention unit of a target type in an attention unit queue in the third network layer, so as to obtain a weight vector, where the weight vector includes n weight values;

a predicted value determining unit 652, configured to process, by the third network layer, the weight vector, the first output vector, and the second output vector to obtain a predicted value of each element to be recommended in the set of elements to be recommended;

the attention unit queue comprises a plurality of attention units of different types, and the target type is matched with the type of the element set to be recommended.

In some possible embodiments, the weight vector determination subunit 651 includes:

a process parameter vector determining subunit 6511, configured to process the first output vector and the second output vector through an attention unit of a target type in an attention unit queue in the third network layer to obtain a process parameter vector, where the process parameter vector includes n process parameters;

a weight vector generation subunit 6512, configured to determine a weight vector according to the process parameter vector, where the weight vector includes n weight values.

In some possible embodiments, the predicted value determining unit 652 includes:

a predicted feature vector determining subunit 6521, configured to process the weight vector, the first output vector, and the second output vector through a splicing collocation network of the third network layer to obtain a predicted feature vector;

a predicted value determining subunit 6522, configured to process the predicted feature vector through a third subnetwork of a target type in a third network in the third network layer, and determine a predicted value corresponding to each element to be recommended;

the third network comprises a plurality of third sub-networks of different types, and the target type is matched with the type of the element set to be recommended.

In some possible embodiments, the fused feature data of each element to be recommended in the set of elements to be recommended includes: and recommending page feature data of the page by the target, wherein the target type is determined according to the page feature data.

Based on the same inventive concept, the principle and the advantageous effect of the problem solving of the recommendation processing apparatus provided in the embodiment of the present application are similar to the principle and the advantageous effect of the problem solving of the message processing method in the embodiment of the present application, and reference may be made to the principle and the advantageous effect of the implementation of the method, and further, the description of the related content in the foregoing embodiment may be referred to for the relationship between the steps executed by the related modules, which is not described herein again for brevity.

Please refer to fig. 8, fig. 8 is a schematic structural diagram of an intelligent device according to an embodiment of the present application. The intelligent device is provided with a ranking model obtained through pre-training, and the ranking model comprises a first network layer, a second network layer and a third network layer. As shown in fig. 8, the intelligent device in this embodiment may include: one or more processors 801, memory 802, and a transceiver 803. The processor 801, the memory 802, and the transceiver 803 are connected by a bus 804.

The processor 801 (or Central Processing Unit (CPU)) is a computing core and a control core of the data Processing device, and can analyze various instructions in the terminal device and process various data of the terminal device, for example: the CPU can be used for analyzing a power-on and power-off instruction sent to the terminal equipment by a user and controlling the terminal equipment to carry out power-on and power-off operation; the following steps are repeated: the CPU may transmit various types of interactive data between the internal structures of the terminal device, and so on. The processor 801 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a Programmable Logic Device (PLD), or the like. The PLD may be a field-programmable gate array (FPGA), a General Array Logic (GAL), or the like.

The transceiver 803 may optionally include a standard wired interface, a wireless interface (e.g., WI-FI, mobile communications interface, etc.), which may be controlled by the processor 801 for transceiving data; the transceiver 803 can also be used for transmission and interaction of data within the terminal device.

The Memory 802(Memory) is a Memory device in the terminal device for storing programs and data. It is understood that the memory 802 may include a built-in memory of the terminal device, and certainly, an expansion memory supported by the terminal device. Memory 802 provides storage space that stores the operating system of the terminal device, which may include, but is not limited to: android system, iOS system, Windows Phone system, etc., which are not limited in this application. The memory 802 may include volatile memory (volatile memory), such as random-access memory (RAM); the memory 802 may also include a non-volatile memory (non-volatile memory), such as a flash memory (flash memory), a solid-state drive (SSD), etc.; the memory 803 may also comprise a combination of memories of the kind described above.

The memory 802 is used to store a computer program comprising program instructions, and the processor 801 and transceiver 803 are used to execute the program instructions stored by the memory 802 to perform the following operations:

In some possible implementations, the processor 801 is configured to:

obtaining the association characteristics of the element set to be recommended, wherein the association characteristics comprise: any one or more of user characteristic data of a target recommending user, page characteristic data of a target recommending page corresponding to an element set to be recommended, and article characteristic data of each element to be recommended in at least two elements to be recommended included in the target recommending page;

and generating fusion characteristic data corresponding to each element to be recommended according to the association characteristics of the element set to be recommended by the first network layer.

In some possible embodiments, the dimension of the fused feature data is [1 × N ]; the processor 801 is configured to:

acquiring a first preset vector, and generating a dense vector according to the fused feature data and the first preset vector, wherein the dimensionality of the first preset vector is [ M, N ], and the dimensionality of the dense vector is [ N, M ];

and determining a first feature vector and a second feature vector corresponding to each element to be recommended according to the dense vector, wherein the dimension of the first feature vector is [1, M ], and the dimension of the second feature vector is [1, N M ].

In some possible embodiments, the second network in the second network layer includes n second subnetworks, n being an integer greater than 1; the processor 801 is configured to:

determining the second feature vector as n vector sets;

and respectively inputting the n vector sets into the n second sub-networks, and respectively processing the corresponding vector sets through the second sub-networks to obtain second output vectors.

In some possible implementations, the processor 801 is configured to:

processing the first output vector and the second output vector through an attention unit of a target type in an attention unit queue in the third network layer to obtain a weight vector, wherein the weight vector comprises n weight values;

processing the weight vector, the first output vector and the second output vector through the third network layer to obtain a predicted value of each element to be recommended in the element set to be recommended;

In some possible implementations, the processor 801 is configured to:

processing the first output vector and the second output vector through an attention unit of a target type in an attention unit queue in the third network layer to obtain a process parameter vector, wherein the process parameter vector comprises n process parameters;

determining a weight vector according to the process parameter vector, wherein the weight vector comprises n weight values.

In some possible implementations, the processor 801 is configured to:

processing the weight vector, the first output vector and the second output vector through a splicing collocation network of the third network layer to obtain a prediction characteristic vector;

processing the predicted feature vector through a third sub-network of a target type in a third network in the third network layer, and determining a predicted value corresponding to each element to be recommended;

It should be appreciated that in some possible implementations, the processor 801 may be a Central Processing Unit (CPU), and the processor may be other general purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and the like. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The memory 802 may include both read-only memory and random access memory, and provides instructions and data to the processor 801. A portion of the memory 802 may also include non-volatile random access memory. For example, the memory 802 may also store device type information.

Based on the same inventive concept, the principle and the beneficial effect of solving the problem of the intelligent device provided in the embodiment of the present application are similar to the principle and the beneficial effect of solving the problem of the message processing method in the embodiment of the present application, and reference may be made to the principle and the beneficial effect of the implementation of the method, and the relationship between the steps executed by the processor 801 may also refer to the description of the related contents in the foregoing embodiment, which is not repeated herein for brevity.

The embodiment of the present application further provides a computer-readable storage medium, where one or more instructions are stored in the computer-readable storage medium, and the one or more instructions are adapted to be loaded by a processor and execute the recommendation processing method described in the foregoing method embodiment.

Embodiments of the present application further provide a computer program product containing instructions, which when run on a computer, cause the computer to execute the recommendation processing method described in the above method embodiments.

Embodiments of the present application also provide a computer program product or a computer program comprising computer instructions stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to execute the method of recommendation processing.

The steps in the method of the embodiment of the application can be sequentially adjusted, combined and deleted according to actual needs.

The modules in the device can be merged, divided and deleted according to actual needs.

Those skilled in the art will appreciate that all or part of the steps in the methods of the above embodiments may be implemented by hardware related to instructions of a program, which may be stored in a computer readable storage medium, which may include: flash disks, Read-Only memories (ROMs), Random Access Memories (RAMs), magnetic or optical disks, and the like.

While the invention has been described with reference to a preferred embodiment, it will be understood by those skilled in the art that various changes in form and detail may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims

1. A recommendation processing method is applied to an intelligent device, a ranking model obtained through pre-training is configured on the intelligent device, the ranking model comprises a first network layer, a second network layer and a third network layer, and the method comprises the following steps:

2. The method according to claim 1, wherein the obtaining of the fused feature data of each element to be recommended in the set of elements to be recommended comprises:

3. The method according to claim 1, wherein the fused feature data has a dimension of [1 x N ]; the processing, by the first network layer, the fused feature data corresponding to each element to be recommended to generate a first feature vector and a second feature vector corresponding to each element to be recommended includes:

4. The method of claim 1, wherein the second network in the second network layer comprises n second subnetworks, n being an integer greater than 1; the processing the second eigenvector corresponding to each element to be recommended through a second network in a second network layer to obtain a second output vector includes:

determining the second feature vector as n vector sets;

5. The method according to any one of claims 1 to 4, wherein the processing the first output vector and the second output vector by a third network layer to obtain a predicted value of each element to be recommended in the set of elements to be recommended includes:

6. The method of claim 5, wherein processing the first output vector and the second output vector by attention units of a target type in an attention unit queue in the third network layer to obtain a weight vector comprises:

7. The method according to claim 5, wherein the processing, by the third network layer, the weight vector, the first output vector, and the second output vector to obtain a predicted value of each element to be recommended in the set of elements to be recommended comprises:

8. The method according to claim 1, 5 or 7, wherein the fused feature data of each element to be recommended in the set of elements to be recommended comprises: and recommending page feature data of the page by the target, wherein the target type is determined according to the page feature data.

9. A recommendation processing apparatus is a network device, a ranking model obtained by pre-training is configured on the apparatus, the ranking model includes a first network layer, a second network layer and a third network layer, and the apparatus includes:

a first output vector determining module, configured to process, through a first sub-network of a target type included in a first network in a second network layer, a first feature vector corresponding to each element to be recommended to obtain a first output vector, where the first network includes first sub-networks of multiple types, and the target type is matched with a type to which the set of elements to be recommended belongs;

10. A smart device comprising a processor, a memory, and a transceiver, the processor, the memory, and the transceiver being interconnected;

the memory for storing a computer program comprising program instructions, the processor and the transceiver being configured to invoke the program instructions to perform the method of any of claims 1-8.

11. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program comprising program instructions that, when executed by a processor, cause the processor to carry out the method according to any one of claims 1-8.