CN114547475B

CN114547475B - Resource recommendation method, device and system

Info

Publication number: CN114547475B
Application number: CN202210448010.6A
Authority: CN
Inventors: 肖钢; 潘建东; 文博; 刘建阳; 张健; 徐政钧
Original assignee: China Securities Co Ltd
Current assignee: China Securities Co Ltd
Priority date: 2022-04-27
Filing date: 2022-04-27
Publication date: 2022-07-26
Anticipated expiration: 2042-04-27
Also published as: CN114547475A

Abstract

The embodiment of the invention provides a resource recommendation method, a device and a system, which are used for acquiring first personal basic information and first resource demand information of a first user; acquiring resource characteristics of each first resource to be recommended and a plurality of different historical recommendation information; processing the first personal basic information, the first resource demand information and the combined information to be recommended based on a pre-trained matching degree prediction model aiming at each combined information to be recommended of the first resource to be recommended to obtain the matching degree of the first user and the combined information to be recommended; the combined information to be recommended comprises the resource characteristics of the first resource to be recommended and historical recommendation information; the historical recommendation information contained in each two pieces of combined information to be recommended is different; determining a first alternative resource of which the matching degree corresponding to the combined information to be recommended meets a preset condition from each first resource to be recommended; a target resource is recommended to the first user that includes the first alternative resource. Based on this, the effectiveness of the recommended resources can be improved.

Description

Resource recommendation method, device and system

Technical Field

The invention relates to the technical field of artificial intelligence, in particular to a resource recommendation method, device and system.

Background

With the development of computer technology, more and more resources can be obtained by users, and it is difficult for users to quickly select interested resources from massive resources (such as movies, music, financial products, etc.).

In the related art, resources may be recommended to a user, for example, when a financial product is recommended to the user, social data of the user on a social platform (e.g., QQ, WeChat, microblog, etc.) for a period of time (e.g., a week) is obtained, and keywords associated with the financial product are extracted from the social data of the user, for example, if the user is interested in a financial product related to a large amount of financing, keywords of the financial product related to the large amount of financing may be included in the social data of the user. Then, the target financial product associated with the extracted keyword is recommended to the user.

However, the keywords associated with the financial products extracted from the social data of the user may not represent the characteristics of the financial products meeting the real needs of the user, for example, the keywords of the financial products related to the large amount of funds are included in the social data of the user, but the assets of the user may not meet the purchasing standards of the financial products related to the large amount of funds, and the financial products related to the large amount of funds do not meet the personal real information of the user. Therefore, the target financial product determined based on the social data of the user may not meet the real demand of the user, and the target financial product is recommended to the user, which may result in low effectiveness of the recommended resource.

Disclosure of Invention

The embodiment of the invention aims to provide a resource recommendation method, device and system so as to improve the effectiveness of recommended resources. The specific technical scheme is as follows:

in a first aspect, to achieve the above object, an embodiment of the present invention provides a resource recommendation method, where the method is applied to a server, and the method includes: acquiring personal basic information of a first user as first personal basic information, and acquiring resource demand information of the first user as first resource demand information; aiming at each first resource to be recommended, acquiring resource characteristics of the first resource to be recommended and a plurality of different historical recommendation information; wherein, a history recommendation information of the first resource to be recommended is: the method comprises the steps that audio data generated when the first resource to be recommended is recommended to a user in a first historical time period are obtained; processing the first personal basic information, the first resource demand information and the combined information to be recommended based on a pre-trained matching degree prediction model aiming at each combined information to be recommended of the first resource to be recommended to obtain the matching degree of the first user and the combined information to be recommended; the combined information to be recommended comprises the resource characteristics of the first resource to be recommended and historical recommendation information; the historical recommendation information contained in each two pieces of combined information to be recommended is different; the matching degree prediction model is as follows: the method comprises the steps that training is carried out on the basis of sample personal basic information of sample users, sample resource demand information and sample combination information of sample resources; determining a resource with matching degree meeting a preset condition corresponding to the combined information to be recommended from each first resource to be recommended as a first alternative resource; recommending a target resource to the first user; wherein the target resource comprises the first alternative resource.

In a second aspect, to achieve the above object, an embodiment of the present invention provides a resource recommendation apparatus, where the apparatus is applied to a server, and the apparatus includes: the first acquisition module is used for acquiring personal basic information of a first user as first personal basic information and acquiring resource demand information of the first user as first resource demand information; the second obtaining module is used for obtaining the resource characteristics of each first resource to be recommended and a plurality of different historical recommendation information; wherein, a history recommendation information of the first resource to be recommended is: the method comprises the steps that audio data generated when the first resource to be recommended is recommended to a user in a first historical time period are obtained; the first determining module is used for processing the first personal basic information, the first resource demand information and the combined information to be recommended based on a pre-trained matching degree prediction model aiming at each combined information to be recommended of the first resource to be recommended to obtain the matching degree of the first user and the combined information to be recommended; the combined information to be recommended comprises resource characteristics of the first resource to be recommended and historical recommendation information; the historical recommendation information contained in each two pieces of combined information to be recommended is different; the matching degree prediction model is as follows: the method comprises the steps that training is carried out on the basis of sample personal basic information of sample users, sample resource demand information and sample combination information of sample resources; the second determining module is used for determining the resource of which the matching degree corresponding to the combined information to be recommended meets the preset condition from each first resource to be recommended as a first alternative resource; the recommending module is used for recommending the target resource to the first user; wherein the target resource comprises the first alternative resource.

In a third aspect, to achieve the above object, an embodiment of the present invention provides a resource recommendation system, where the system includes: first client, second client and server, wherein:

the first client is used for sending first audio data of a first user for acquiring resources to the server;

the server is used for carrying out voice recognition on the first audio data when the first audio data are received, so as to obtain first resource demand information of the first user; aiming at each first resource to be recommended, acquiring resource characteristics of the first resource to be recommended and a plurality of different historical recommendation information; wherein, a history recommendation information of the first resource to be recommended is: the method comprises the steps that audio data generated when the first resource to be recommended is recommended to a user in a first historical time period are obtained; processing the first personal basic information of the first user, the first resource demand information and the combined information to be recommended based on a pre-trained matching degree prediction model aiming at each combined information to be recommended of the first resource to be recommended to obtain the matching degree of the first user and the combined information to be recommended; the combined information to be recommended comprises the resource characteristics of the first resource to be recommended and historical recommendation information; the historical recommendation information contained in each two pieces of combined information to be recommended is different; the matching degree prediction model is as follows: the method comprises the steps that training is carried out on the basis of sample personal basic information of sample users, sample resource demand information and sample combination information of sample resources; determining a resource with matching degree meeting a preset condition corresponding to the combined information to be recommended from each first resource to be recommended as a first alternative resource; target recommendation information of the target resource is sent to the second client logged in by the second user; wherein the target resource comprises the first alternative resource; the target recommendation information of the target resource comprises: historical recommendation information in the combined information of which the corresponding matching degree of the target resource meets the preset conditions;

the second client is used for displaying the received target recommendation information, acquiring second audio data sent by the second user according to the target recommendation information, and sending the second audio data to the server;

the server is further used for sending the second audio data to the first client when receiving the second audio data;

the first client is further configured to play the second audio data.

The embodiment of the invention also provides electronic equipment which comprises a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory are communicated with each other through the communication bus; a memory for storing a computer program; and the processor is used for realizing any one of the steps of the resource recommendation method when the program stored in the memory is executed.

An embodiment of the present invention further provides a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, the steps of any of the resource recommendation methods described above are implemented.

Embodiments of the present invention further provide a computer program product containing instructions, which when run on a computer, cause the computer to execute any of the above resource recommendation methods.

According to the resource recommendation method provided by the embodiment of the invention, the personal basic information of a first user is acquired as the first personal basic information, and the resource demand information of the first user is acquired as the first resource demand information; aiming at each first resource to be recommended, acquiring resource characteristics of the first resource to be recommended and a plurality of different historical recommendation information; one historical recommendation information of the first resource to be recommended is as follows: the method comprises the steps that audio data generated when the first resource to be recommended is recommended to a user in a first historical time period are obtained; processing the first personal basic information, the first resource demand information and the combined information to be recommended based on a pre-trained matching degree prediction model aiming at each combined information to be recommended of the first resource to be recommended to obtain the matching degree of the first user and the combined information to be recommended; the combined information to be recommended comprises the resource characteristics of the first resource to be recommended and historical recommendation information; the historical recommendation information contained in each two pieces of combined information to be recommended is different; the matching degree prediction model is as follows: the method comprises the steps that training is carried out on the basis of sample personal basic information of sample users, sample resource demand information and sample combination information of sample resources; determining a resource with matching degree meeting a preset condition corresponding to the combined information to be recommended from each first resource to be recommended as a first alternative resource; recommending a target resource to a first user; the target resource includes a first alternative resource.

Based on the processing, the first resource demand information can embody the characteristics of the resource required by the first user, the first personal basic information can embody the real information of the first user, and the to-be-recommended combined information of the first to-be-recommended resource includes the resource characteristics of the first to-be-recommended resource. Correspondingly, the matching degree of the first user and the combined information to be recommended, which is determined based on the first personal basic information, the first resource demand information and the combined information to be recommended, can embody the degree that the first resource to be recommended meets the real demand of the first user. Furthermore, the first alternative resource determined based on the matching degree of the first user and the combined information to be recommended of each first resource to be recommended meets the real requirement of the user, the target resource containing the first alternative resource is recommended to the first user, and the effectiveness of the recommended resource can be improved. In addition, the historical recommendation information contained in each piece of combined information to be recommended of the first resource to be recommended is different, and the matching degree between one piece of combined information to be recommended and the first user can represent the degree that the historical recommendation information in the combined information to be recommended meets the real requirement of the first user. And determining target recommendation information of the target resource according with the real requirement of the first user based on the matching degree of each to-be-recommended combination information and the first user. Target resources are recommended to the first user based on the target recommendation information, personalized resource recommendation can be performed on the basis of the target recommendation information which meets the real requirements of the users aiming at different users, and the effectiveness of the recommended resources is further improved.

Of course, it is not necessary for any product or method to achieve all of the above-described advantages at the same time for practicing the invention.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art that other embodiments can be obtained by referring to these drawings.

Fig. 1 is a flowchart of a resource recommendation method according to an embodiment of the present invention;

FIG. 2 is a flowchart of another resource recommendation method according to an embodiment of the present invention;

FIG. 3 is a flowchart of another resource recommendation method according to an embodiment of the present invention;

FIG. 4 is a flowchart of another resource recommendation method according to an embodiment of the present invention;

FIG. 5 is a flowchart of a model training method according to an embodiment of the present invention;

FIG. 6 is a flow chart of a resource recommendation method in the prior art;

FIG. 7 is a block diagram of a resource recommendation system according to an embodiment of the present invention;

FIG. 8 is a block diagram of another resource recommendation system provided in an embodiment of the present invention;

FIG. 9 is a flowchart of another resource recommendation method according to an embodiment of the present invention;

FIG. 10 is a schematic diagram of a resource recommendation method according to an embodiment of the present invention;

fig. 11 is a flowchart of a training sample generation method according to an embodiment of the present invention;

FIG. 12 is a flowchart of another training sample generation method according to an embodiment of the present invention;

fig. 13 is a structural diagram of a resource recommendation device according to an embodiment of the present invention;

fig. 14 is a structural diagram of an electronic device according to an embodiment of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived from the embodiments given herein by one of ordinary skill in the art, are within the scope of the invention.

Referring to fig. 1, fig. 1 is a flowchart of a resource recommendation method provided by an embodiment of the present invention, where the method is applied to a server, and the method includes the following steps:

s101: the method comprises the steps of obtaining personal basic information of a first user as first personal basic information, and obtaining resource demand information of the first user as first resource demand information.

S102: and for each first resource to be recommended, acquiring the resource characteristics of the first resource to be recommended and a plurality of different historical recommendation information. The historical recommendation information of the first resource to be recommended is as follows: the recommendation method comprises the steps of obtaining audio data generated when the first resource to be recommended is recommended to a user in a first historical time period.

S103: and processing the first personal basic information, the first resource demand information and the combined information to be recommended based on a pre-trained matching degree prediction model aiming at each combined information to be recommended of the first resource to be recommended to obtain the matching degree of the first user and the combined information to be recommended. And the combined information to be recommended comprises the resource characteristics of the first resource to be recommended and historical recommendation information. And the historical recommendation information contained in each two pieces of combined information to be recommended is different. The matching degree prediction model is as follows: the method is obtained by training based on sample personal basic information of sample users, sample resource demand information and sample combination information of sample resources.

S104: and determining the resource of which the matching degree corresponding to the combined information to be recommended meets the preset condition from the first resources to be recommended as the first alternative resource.

S105: the target resource is recommended to the first user. Wherein the target resource comprises a first alternative resource.

Based on the resource recommendation method provided by the embodiment of the invention, the first resource demand information can embody the characteristics of the resources required by the first user, the first personal basic information can embody the real information of the first user, and the to-be-recommended combined information of the first to-be-recommended resources includes the resource characteristics of the first to-be-recommended resources. Correspondingly, the matching degree of the first user and the combined information to be recommended, which is determined based on the first personal basic information, the first resource demand information and the combined information to be recommended, can embody the degree that the first resource to be recommended meets the real demand of the first user. Furthermore, the first alternative resource determined based on the matching degree of the first user and the combined information to be recommended of each first resource to be recommended meets the real requirement of the user, the target resource containing the first alternative resource is recommended to the first user, and the effectiveness of the recommended resource can be improved. In addition, the historical recommendation information contained in each piece of combined information to be recommended of the first resource to be recommended is different, and the matching degree between one piece of combined information to be recommended and the first user can represent the degree that the historical recommendation information in the combined information to be recommended meets the real requirement of the first user. And determining target recommendation information of the target resource according with the real requirement of the first user based on the matching degree of each to-be-recommended combination information and the first user. Target resources are recommended to the first user based on the target recommendation information, personalized resource recommendation can be performed on the basis of the target recommendation information which meets the real requirements of the users aiming at different users, and the effectiveness of the recommended resources is further improved.

With respect to step S101, the first resource to be recommended is a resource that has been recommended to the user among all resources provided by the server to the user. When the method provided by the embodiment of the invention is applied to the financial industry, the first resource to be recommended can be a financial product, such as fund, stock, periodic deposit, national debt and the like. The first user may be any user that currently needs to make a resource recommendation. The first personal basic information of the first user represents user information of the first user. The first personal basic information may include at least one of: revenue information of the first user (e.g., annual revenue of the first user), total asset information (e.g., deposit amount of the first user), age, preferred financial risk level, information of financial products historically traded (e.g., financial products bought and financial products sold by the first user within the past month), and the like.

The first resource requirement information of the first user represents: the first user needs to obtain characteristics of the resource. For example, when the first resource to be recommended is a financial product, the first resource demand information may include: the first user needs to acquire the annual rate of return, the minimum amount of purchase, the minimum time of purchase, and the like of the financial product.

In an embodiment of the present invention, a resource recommendation may be performed by a second user to a first user in a manner that a first client where the first user logs in performs real-time voice interaction with a second client where the second user logs in. Accordingly, the server may obtain the first resource requirement information based on the voice data of the first user. The second user is a staff member who performs resource recommendation to the first user, for example, when the first user needs a financial product recommendation, the second user may be a bank staff member or a stock company staff member.

Correspondingly, the process of acquiring the first resource requirement information by the server in step S101 may include the following steps:

step 1, receiving first audio data which is sent by a first client terminal logged in by a first user and used for acquiring resources.

And 2, performing voice recognition on the first audio data to obtain resource demand information of the first user as the first resource demand information.

When a first user needs to acquire a resource, the first user inputs a resource acquisition instruction to a first client. And when receiving the resource acquisition instruction, the first client sends a voice communication request aiming at the second client to the server. And when receiving the voice communication request, the server establishes communication connection between the first client and the second client. Subsequently, the first client and the second client perform real-time voice interaction through the server so as to realize that the second user performs resource recommendation to the first user in real time. Furthermore, the first user describes the characteristics of the resource that needs to be acquired through voice, and inputs audio data (i.e., first audio data) for acquiring the resource to the first client, where the first audio data includes the characteristics of the resource that the first user needs to acquire. The first client may send the first audio data to the server when acquiring the first audio data. When the server receives the first audio data, the server can perform voice recognition on the first audio data to extract the characteristics of the resource required to be acquired by the first user, so as to obtain first resource demand information.

For example, the server may perform text conversion on the first audio data to obtain a corresponding text, where the text includes the first resource requirement information. The server may text-convert the first audio data based on an ASR (Automatic Speech Recognition) algorithm. Or the first audio data may be text-converted by PocketSphinx (an open source tool for speech to text). Subsequently, the server may determine, based on the obtained first personal basic information and the first resource demand information, a target resource recommended to the first user based on the method provided by the embodiment of the present invention, and perform resource recommendation to the first user through the second client.

The personal basic information of the user is stored in a preset database in advance, and the server can obtain the personal basic information corresponding to the user identification of the first user from the preset database to obtain the first personal basic information.

In an embodiment of the present invention, in order to ensure information security of the user, when acquiring the first personal basic information, the server may send a reminding message to the first client for reminding the first user whether to authorize the server to acquire the first personal basic information. The first client may display the received reminder message to remind the first user whether to authorize the server to obtain the first personal basic information. And when receiving an authorization confirmation instruction input by the first user, the first client sends the authorization confirmation instruction to the server. And when receiving the authorization confirmation instruction, the server acquires the first personal basic information. Based on the processing, the personal basic information of the user can be obtained under the condition that the user is authorized, the information safety of the user can be protected, and the relevant legal regulations are met.

For step S102, a resource feature of a first resource to be recommended represents resource information of the first resource to be recommended. The resource characteristics of a first resource to be recommended include at least one of: the name, code, position taken, historical rate of return (e.g., the rate of return per day of the first resource to be recommended in the past month), manager information, belonging business segment, financial risk level, minimum purchase amount, minimum purchase time, and the like of the first resource to be recommended.

The historical recommendation information of a first resource to be recommended is as follows: and recommending sentences used when the first resource to be recommended is recommended to the user. The historical recommendation information of a first resource to be recommended is as follows: the recommendation method is based on audio data (which may be called historical audio data) generated when the first resource to be recommended is recommended to the user in the first historical time period. The duration of the first historical time period may be set based on the demand, and when more historical recommendation information of the first resource to be recommended needs to be acquired, the first historical time period may be set to be longer, for example, one year, six months, and the like. When less historical recommendation information of the first resource to be recommended needs to be acquired, the first historical time period may be set to be shorter, for example, one month, one week, or the like.

The historical audio data corresponding to the first resource to be recommended comprises: and/or the audio data is generated when the first resource to be recommended is recommended to the user through the second client in the first history time period. The server may perform voice recognition (e.g., text conversion) on the historical audio data corresponding to the first resource to be recommended, so as to obtain historical recommendation information of the first resource to be recommended.

For step S103, for each first resource to be recommended, in the first history time period, different recommendation information (i.e., history recommendation information) may be used each time the first resource to be recommended is recommended to the user, and the history recommendation information of the first resource to be recommended is multiple. The combined information to be recommended of a first resource to be recommended comprises the resource characteristics of the first resource to be recommended and historical recommendation information; and the historical recommendation information contained in each two pieces of combined information to be recommended is different. Illustratively, the historical recommendation information of the first resource to be recommended includes: recommending information 1, recommending information 2, recommending information 3 and recommending information 4, wherein the combined information to be recommended of the first resource to be recommended comprises: combination information 1 including recommendation information 1 and resource features, combination information 2 including recommendation information 2 and resource features, combination information 3 including recommendation information 3 and resource features, and combination information 4 including recommendation information 4 and resource features.

For each first resource to be recommended, the historical recommendation information contained in each piece of combined information to be recommended of the first resource to be recommended is different, and the matching degree of one piece of combined information to be recommended with the first user can represent the degree that the historical recommendation information in the combined information to be recommended meets the real requirement of the first user. Correspondingly, the determined target recommendation information of the target resource meets the real requirement of the first user based on the matching degree of each combination information to be recommended and the first user. Subsequently, the target resource is recommended to the first user based on the target recommendation information, personalized resource recommendation can be performed on the basis of the target recommendation information meeting the real requirements of the users aiming at different users, and the effectiveness of the recommended resource is improved. For example, when a technology-based fund is recommended to a user (e.g., a programmer) in a technology-based professional background, the user knows technology-based industry information associated with the technology-based fund, and the target recommendation information may include fewer descriptions of the technology-based industry information, so as to improve the communication efficiency with the user and further improve the resource recommendation efficiency. When recommending to a user with a non-technical professional background, the user does not know the scientific and technological industry information associated with the scientific and technological fund, and the target recommendation information can contain more scientific and technological industry information introductions, so that the user can further know the relevant information of the scientific and technological fund recommended to the user, and the effectiveness of the recommended resources is improved.

The matching degree prediction model may be a double tower model, for example, a DSSM (Deep Structured Semantic model), or an NCF (Neural Collaborative Filtering) model, which is not limited in the embodiments of the present invention.

In one embodiment of the present invention, the matching degree prediction model includes: the system comprises a resource feature processing module, a user feature processing module and a matching degree calculating module. The resource characteristic processing module comprises: the system comprises a first feature mapping module, a first feature fusion module and a first Multilayer Perceptron (MLP for short). The user characteristic processing module comprises: the system comprises a second feature mapping module, a second feature fusion module and a second multilayer perceptron.

Accordingly, on the basis of fig. 1, referring to fig. 2, step S103 may include the following steps:

s1031: and for each piece of combined information to be recommended of the first resource to be recommended, mapping the resource features in the combined information to be recommended through a first feature mapping module to obtain a first feature vector, and mapping historical recommendation information in the combined information to be recommended to obtain a second feature vector.

S1032: and splicing the first feature vector and the second feature vector through a first feature fusion module to obtain a third feature vector.

S1033: and mapping the third characteristic vector according to the specified length through the first multilayer perceptron to obtain the information characteristic vector of the combined information to be recommended.

S1034: and mapping the first personal basic information through a second feature mapping module to obtain a fourth feature vector, and mapping the first resource demand information to obtain a fifth feature vector.

S1035: and splicing the fourth feature vector and the fifth feature vector through a second feature fusion module to obtain a sixth feature vector.

S1036: and mapping the sixth feature vector according to the specified length through the second multilayer perceptron to obtain the user feature vector of the first user.

S1037: and calculating the similarity between the user characteristic vector of the first user and the information characteristic vector of the combined information to be recommended through a matching degree calculation module to serve as the matching degree of the first user and the combined information to be recommended.

The mapping processing mode of the first feature mapping module may be Onehot encoding or Embedding encoding.

In an implementation manner, for each piece of to-be-recommended combination information of each first to-be-recommended resource, the server may perform Onehot coding on a non-sequence type resource feature (for example, a name of the first to-be-recommended resource) in the piece of to-be-recommended combination information through the first feature mapping module, and perform embedded coding on a sequence type resource feature (for example, a rate of return of the first to-be-recommended resource every day in the past month) in the piece of to-be-recommended combination information through the first feature mapping module to obtain the first feature vector. The server can perform Embedded coding on the historical recommendation information in the combined information to be recommended through the first feature mapping module to obtain a second feature vector. Then, the first feature vector and the second feature vector are input into the first feature fusion module, the server can determine the preset weight of the first feature vector and the preset weight of the second feature vector, and the first feature fusion module splices the first feature vector and the second feature vector according to the corresponding preset weights to obtain a third feature vector. Furthermore, the third feature vector may be input to the first multilayer sensor, and the server may map the third feature vector to a specified length through the first multilayer sensor, so as to implement dimension reduction on the third feature vector, and obtain the information feature vector of the combined information to be recommended.

For each piece of combined information to be recommended, when the combined information to be recommended includes a sequence-type resource feature, a third feature vector is determined based on the sequence-type resource feature, the third feature vector includes a sequence-type feature, and in order to process the sequence-type feature, the first multilayer sensor may be any one of the following models: convolutional neural network models, cyclic neural network models, self-attention network models, and the like. For example, the first multi-layer perceptron may be an LSTM (Long Short Term Memory) network model.

The mapping processing mode of the second feature mapping module can be Onehot coding or Embedding coding.

In one implementation, the server may perform Onehot encoding on the first personal basic information (e.g., the age of the first user) of the non-serial type through the second feature mapping module, and perform Embedding encoding on the first personal basic information (e.g., the information of the financial product that the first user transacts in the past month) of the serial type through the second feature mapping module to obtain the fourth feature vector. The information of the financial products that the first user has transacted within the past month may include: financial products purchased, sold, and price variations of financial products by the first user in the past month. The server may perform Embedding coding on the first resource demand information through the second feature mapping module to obtain a fifth feature vector. Then, the fourth feature vector and the fifth feature vector are input into the second feature fusion module, the server can determine the preset weight of the fourth feature vector and the preset weight of the fifth feature vector, and the fourth feature vector and the fifth feature vector are spliced through the second feature fusion module according to the corresponding preset weights to obtain a sixth feature vector. Further, a sixth feature vector may be input to the second multi-layered perceptron. The server may map the sixth feature vector to a specified length through the second multilayer perceptron, so as to implement dimension reduction on the sixth feature vector, and obtain the user feature vector of the first user.

When the sequence-type personal basic information is not included in the first personal basic information or is less included in the first personal basic information, the second multi-layered sensor may include a full link layer. For example, the second multi-layered perceptron may contain three fully-connected layers using ReLU as an activation function. For each piece of combined information to be recommended of each first resource to be recommended, the information feature vector of the combined information to be recommended and the user feature vector of the first user can be input to the matching degree calculation module, and the server can calculate the similarity between the information feature vector of the combined information to be recommended and the user feature vector of the first user through the matching degree calculation module based on a budget similarity algorithm, and the similarity serves as the matching degree between the first user and the combined information to be recommended. The preset similarity algorithm can be a cosine similarity algorithm, an Euclidean distance algorithm, a Chebyshev distance algorithm and the like.

In an embodiment of the present invention, for each to-be-recommended combined information of each first to-be-recommended resource, since the resource feature and the historical recommended information in the to-be-recommended combined information are fixed and unchangeable within a period of time, correspondingly, the information feature vector of the to-be-recommended combined information is also fixed and unchangeable within a period of time. Therefore, after the information feature vector of the combined information to be recommended is determined, the server can also store the information feature vector of the combined information to be recommended. For example, the information feature vector of the combined information to be recommended is stored locally in the server, or the information feature vector of the combined information to be recommended is stored in a preset database. Subsequently, when resource recommendation is performed on other users, the information feature vector of each piece of combined information to be recommended can be directly obtained, resource recommendation is performed on the users based on the obtained information feature vector of each piece of combined information to be recommended, the information feature vector of the combined information to be recommended of the first resource to be recommended does not need to be calculated again, the calculation amount can be reduced, and the resource recommendation efficiency is improved.

In step S104, after determining the matching degree between the first user and the to-be-recommended combination information of each first to-be-recommended resource, the server may determine a first candidate resource recommended to the first user from the first to-be-recommended resources in the following manner.

In the first embodiment, the preset condition may be: the degree of match with the first user is greater than a threshold degree of match. The server can determine the combined information to be recommended, of which the matching degree with the first user is greater than the threshold value of the matching degree, from the plurality of pieces of combined information to be recommended of the plurality of first resources to be recommended, as the target combined information. Then, the server may determine a first resource to be recommended corresponding to the target combination information as a first candidate resource. The threshold of the matching degree may be set by a skilled person based on experience, for example, the threshold of the matching degree may be 0.8, or may also be 0.7, but is not limited thereto.

In a second mode, the preset condition may be: the matching degree with the first user is larger than the matching degree of other combined information to be recommended with the first user. The server may determine, in order from a large matching degree with the first user to a small matching degree, a second number of previous pieces of combined information to be recommended from the pieces of combined information to be recommended of the first pieces of resources to be recommended, as the target combined information. Then, the server may determine the first resource to be recommended corresponding to the target combination information as the first candidate resource.

In an embodiment of the present invention, on the basis of fig. 1, referring to fig. 3, before step S105, the method may further include the steps of:

s106: and determining a plurality of resources without the historical recommendation information as second resources to be recommended.

S107: and for each second resource to be recommended, mapping the resource characteristics of the second resource to be recommended and preset recommendation information to obtain a resource characteristic vector of the second resource to be recommended.

S108: and calculating the similarity between the resource feature vector of the second resource to be recommended and the information feature vector of the target combination information of each first alternative resource to obtain the similarity between the second resource to be recommended and the first alternative resource. And the matching degree corresponding to the target combination information of the first alternative resource meets a preset condition.

S109: and calculating the mean value of the similarity of the second resource to be recommended and each first alternative resource to obtain the mean value of the similarity corresponding to the second resource to be recommended.

S1010: and determining the first number of resources from the second resources to be recommended as second alternative resources according to the sequence of the corresponding similarity mean values from large to small. Wherein the target resource further comprises a second alternative resource.

The second resource to be recommended is a new resource which is not recommended to the user in all resources provided by the server to the user. Because the second resource to be recommended is not recommended to the user, the second resource to be recommended cannot be recommended to the user based on the historical recommendation information of the second resource to be recommended if the second resource to be recommended does not have the historical recommendation information. In order to recommend the second resource to be recommended to the user, for each second resource to be recommended, the server may perform mapping processing on the resource features of the second resource to be recommended to obtain a corresponding feature vector, and perform mapping processing on the preset recommendation information to obtain a corresponding feature vector. And then splicing the two obtained feature vectors to obtain the resource feature vector of the second resource to be recommended. The preset recommendation information of the second resource to be recommended is preset by a technician.

Then, the server may calculate a similarity between the resource feature vector of the second resource to be recommended and the information feature vector of the target combination information of each first candidate resource, to obtain a similarity between the second resource to be recommended and the first candidate resource. The target combination information of the first alternative resource is: and the matching degree corresponding to the first alternative resource meets the combined information to be recommended of the preset condition. And then, calculating the mean value of the similarity between the second resource to be recommended and each first candidate resource to obtain the mean value of the similarity corresponding to the second resource to be recommended, determining the first number of resources from the second resource to be recommended as the second candidate resource according to the descending order of the corresponding mean values of the similarity, and determining the second candidate resource as the target resource. The first number may be determined based on a second number of the first alternative resources. For example, the first number may be the second number plus one.

In addition, the server may also rank the target resources, that is, rank the first alternative resource and the second alternative resource. Subsequently, the target resources can be recommended to the first user in sequence according to the arrangement sequence of the target resources.

In one implementation manner, the server may sort the first alternative resources according to a descending order of matching degrees corresponding to the target combination information of the first alternative resources, and sort the second alternative resources according to a descending order of similarity mean values corresponding to the second alternative resources. And determining the arrangement sequence of the target resources based on the arrangement sequence of the first alternative resources and the arrangement sequence of the second alternative resources. Illustratively, the first candidate resources are ranked according to the sequence from the largest matching degree to the smallest matching degree corresponding to the target combination information of the first candidate resources, so as to obtain: resource 1, resource 2, resource 3, and resource 4. And sequencing the second alternative resources according to the descending order of the similarity mean value corresponding to the second alternative resources to obtain: resource 5, resource 6, resource 7, resource 8, and resource 9. Then, the server may arrange the second candidate resources in the order, and arrange one second candidate resource before one first candidate resource, and the arrangement order of the target resources may be obtained as follows: resource 5, resource 1, resource 6, resource 2, resource 7, resource 3, resource 8, resource 4, and resource 9.

For step S105, the target resource may include a first candidate resource determined from the first resource to be recommended, or the target resource may also include a first candidate resource determined from the first resource to be recommended and a second candidate resource determined from the second resource to be recommended.

In an embodiment of the present invention, on the basis of fig. 1, referring to fig. 4, step S105 may include the following steps:

s1051: and sending the target recommendation information of the target resource to a second client logged in by a second user so as to enable the second client to display the received target recommendation information, acquire second audio data sent by the second user according to the target recommendation information, and send the second audio data to the server. The target recommendation information of the target resource comprises: and the corresponding matching degree of the target resource meets the historical recommendation information in the combined information of the preset conditions.

S1052: and when the second audio data is received, sending the second audio data to the first client so that the first client plays the second audio data.

For a scene that a first client and a second client perform real-time voice interaction to recommend resources to a first user, after determining target resources, a server can send target recommendation information of the target resources to the second client. And aiming at each target resource, if the target resource is the second alternative resource, the target recommendation information of the target resource is preset recommendation information. And if the target resource is the first alternative resource, the target recommendation information of the target resource is historical recommendation information in the target combination information, wherein the corresponding matching degree meets the preset condition. Correspondingly, the second client may display the received target recommendation information for the second user to browse, and the second client may obtain audio data (i.e., second audio data) sent when the second user reads the target recommendation information and send the second audio data to the server. The server may send the second audio data to the first client upon receiving the second audio data. The first client can play the received second audio data, so that the second user can recommend the target resource to the first user according to the target recommendation information in a real-time voice interaction mode with the second client through the first client.

Based on the processing, the first resource demand information can reflect the demand of the first user, the first personal basic information can reflect the real information of the first user, resource recommendation is carried out based on the first resource demand information and the first personal basic information, the real information and the demand of the first user can be comprehensively considered, resource recommendation can be carried out on the user more flexibly, and the effectiveness of resource recommendation is improved. For example, the financial risk level matching the user determined based on the user's real information is medium risk, but the user's needs are: knowing the financial products with low risk level can recommend the financial products with low risk level to the user when recommending the financial products by combining the real information and the requirements of the user.

In an embodiment of the present invention, the server may further train the matching degree prediction model of the initial structure to obtain a trained matching degree prediction model. Accordingly, referring to fig. 5, the process of training the initial structure matching degree prediction model may include the following steps:

s501: and for each sample resource, obtaining sample audio data generated when a second user recommends resources to the sample user in a second historical time period, and performing voice recognition on the sample audio data to obtain historical recommendation information of the sample resource and sample resource demand information of the sample user corresponding to the sample resource.

S502: and generating an initial training sample based on sample combination information of the sample resource, sample resource demand information and sample personal basic information of the sample user. Wherein one sample combination information includes: the resource characteristics of the sample resource and historical recommendation information; the historical recommendation information contained in each two sample combination information is different.

S503: and determining a positive sample matched with the sample user and a negative sample not matched with the sample user from the initial training samples based on the sample labels in the initial training samples to obtain the target training sample. Wherein the sample label in an initial training sample represents: the matching degree of the sample combination information in the initial training sample and the sample user.

S504: and adjusting model parameters of the matching degree prediction model of the initial structure based on the target training sample until a preset convergence condition is reached, so as to obtain the trained matching degree prediction model.

The second historical period of time may be set based on demand, and may be the same as the first historical period of time, or may be different from the first historical period of time.

In one implementation, the server may obtain sample audio data for resource recommendation from the second user to the sample user within the second historical time period, and then, the server may perform text conversion on the sample audio data to obtain a corresponding sample text. The sample text comprises sample resource demand information of the sample user and historical recommendation information of the second user when recommending sample resources to the sample user. The server can determine texts containing the recommended keywords from the sample texts as historical recommendation information of the sample resources. The recommendation keywords may include at least one of: suggestions, purchases, recommendations, codes, names of financial products, etc. The server determines the audio corresponding to the historical recommendation information in the sample audio (which may be referred to as sample recommendation audio data), and determines the audio data before the sample recommendation audio data (which may be referred to as sample demand audio data), where the sample demand audio data is sent by the sample user and is located before the sample recommendation audio data, and then the sample demand audio data includes the sample resource demand information corresponding to the sample resource. Therefore, the server can determine a text corresponding to the sample demand audio data from the sample text to obtain sample resource demand information corresponding to the sample resource.

In another implementation manner, the server may obtain sample audio data for resource recommendation to the sample user by the second user in the second historical time period, where the sample audio data includes sample resource demand information of the sample user and historical recommendation information when the sample resource is recommended to the sample user by the second user. For each sample resource, the server may determine, from the sample audio data, sample recommendation audio data issued by the second user for recommending the sample resource. For example, the server may determine audio data containing the recommendation keyword from the sample audio data as the sample recommendation audio data. Then, the server can perform voice recognition on the sample recommended audio data to obtain historical recommendation information of the sample resource. For example, the server may perform text conversion on the sample recommended audio data to obtain the historical recommendation information of the sample resource.

The server may further determine, from the sample audio data, audio data that is sent by the sample user and is located before the sample recommended audio data, and obtain sample demand audio data of the sample user corresponding to the sample resource, where the sample demand audio data is sent by the sample user and is located before the sample recommended audio data, and then the sample demand audio data includes sample resource demand information corresponding to the sample resource. Therefore, the server can perform voice recognition on the sample demand audio data corresponding to the sample resource to obtain the sample resource demand information corresponding to the sample resource. In addition, if the audio data before the sample recommended audio data does not exist in the acquired sample audio data, the server may acquire other audio data before the sample audio data, where the second user recommends the resource to the sample user, and determine the audio data sent by the sample user from the acquired audio data, as the sample required audio data corresponding to the sample resource.

For each sample resource, because the second user recommends the sample resource to different users based on different historical recommendation information, the historical recommendation information of the sample resource is multiple, and the sample resource requirements of different sample users are different, the sample resource requirements corresponding to the sample resource are multiple. Therefore, the server can generate an initial training sample based on sample combination information of the sample resource, sample resource demand information and sample personal basic information of the sample user. One sample combination information includes: the resource characteristics of the sample resource and historical recommendation information; the historical recommendation information contained in each two sample combination information is different. And the sample personal basic information of the sample user is obtained from a preset database under the condition that the sample user is authorized.

Then, the server may determine, based on the matching degree between the sample combination information of the sample resource and the sample user, a sample label corresponding to one piece of sample combination information, one piece of sample resource demand information, and sample personal basic information of the sample user of the sample resource, and obtain an initial training sample including the corresponding sample label. An initial training sample comprising: sample combination information, sample resource requirement information, sample individual basic information and sample labels, wherein the sample labels in an initial training sample represent: the matching degree of the sample combination information in the initial training sample and the sample user. The manner in which the server determines the sample label may refer to the relevant description of the subsequent embodiments.

For example, the historical recommendation information for the sample resource includes: recommendation information 1, recommendation information 2, and recommendation information 3. The sample combination information of the sample resource includes: sample combination information 1 containing resource characteristics of the sample resources and recommendation information 1, sample combination information 2 containing resource characteristics of the sample resources and recommendation information 2, and sample combination information 3 containing resource characteristics of the sample resources and recommendation information 3. The sample resource demand information when recommending the sample resource to the user based on the recommendation information 1 is: resource demand information 1; the sample resource demand information when recommending the sample resource to the user based on the recommendation information 2 is as follows: resource demand information 2; the sample resource demand information when recommending the sample resource to the user based on the recommendation information 3 is: resource demand information 3. The server can take the sample combination information 1, the resource demand information 1, the sample personal basic information and the sample label of the sample resource as an initial training sample 1; taking sample combination information 2, resource demand information 2, sample personal basic information and sample labels of the sample resources as an initial training sample 2; and taking the sample combination information 3, the resource demand information 3, the sample personal basic information and the sample label of the sample resource as an initial training sample 3.

Further, the server may determine, from the initial training samples, a positive sample matching the sample user and a negative sample not matching the sample user to obtain a target training sample, that is, the target training sample includes the positive sample and the negative sample. The sample label of the positive sample matched with the sample user is 1, that is, the matching degree of the sample combination information in the positive sample and the sample user is 1. The sample label of the negative sample that does not match the sample user is 0, that is, the matching degree of the sample combination information in the negative sample and the sample user is 0.

Furthermore, the server may input the sample combination information, the sample resource demand information, and the sample personal basic information in the target training sample into the matching degree prediction model of the initial structure, to obtain an information feature vector (may be referred to as a sample information feature vector) of the sample combination information in the target training sample, and a user feature vector (may be referred to as a sample user feature vector) of the sample user. Then, the matching degree prediction model of the initial structure may calculate a similarity between the sample information feature vector and the sample user feature vector based on the following formula, to obtain a matching degree (which may be referred to as a prediction matching degree) between the sample user and the sample combination information in the target training sample.

（1）

A represents the matching degree of the sample user and the sample combination information in the target training sample, sim () represents a cosine similarity function, u represents a sample information feature vector, and i represents a sample user feature vector.

The server may calculate a loss function value representing the difference between the predicted match degree and a sample match degree, the sample match degree being the match degree represented by the sample label in the target training sample. The loss function may be a cross-entropy loss function as follows:

（2）

l represents a loss function value, u represents a sample information feature vector, i represents a sample user feature vector,

the degree of matching of the samples is represented,

representing the degree of predictive matching.

The server can adjust model parameters of the matching degree prediction model of the initial structure based on the calculated loss function value until a preset convergence condition is reached, and a trained matching degree prediction model is obtained. The preset convergence condition may be that the training times for training the matching degree prediction model of the initial structure reach a preset number, or the preset convergence condition may be that the loss function values obtained by calculating the third number of consecutive times are all smaller than a preset value.

In an embodiment of the present invention, before step S504, the method may further include the steps of: and for each initial training sample, inputting the initial training sample and historical recommended behavior information of sample resources corresponding to the initial training sample into a pre-trained classification model to obtain a sample label of the initial training sample. Wherein the historical recommendation behavior information of a sample resource comprises at least one of: the method comprises the following steps that compliance evaluation information of a sample user for the sample resource, purchasing behavior information of the sample user for the sample resource, historical operation behavior information of the sample user for the sample resource, the behavior information of the sample user for consulting the second user about the sample resource, and satisfaction evaluation information of the sample user for recommending the historical recommending behavior of the sample resource by the second user;

step S504 may include the steps of: step 1: a number of first initial training samples whose sample labels represent matches with the sample user and a number of second initial training samples whose sample labels represent mismatches with the sample user are determined.

Step 2: a ratio of the number of first initial training samples to the number of second initial training samples is calculated.

And step 3: and if the calculated ratio is larger than a preset threshold value, determining the first initial training sample as a positive sample, and determining the second initial training sample as a negative sample to obtain the target training sample.

And 4, step 4: and if the calculated ratio is not larger than the preset threshold value, calculating the similarity between the first initial training sample and each third initial training sample aiming at each first initial training sample. Wherein the third initial training sample is: training samples other than the first initial training sample and the second initial training sample.

And 5: and determining the first initial training sample and a third initial training sample with the maximum similarity to the first initial training sample as positive samples, and determining the second initial training sample as a negative sample to obtain a target training sample.

The Classification model may be a random forest Classification model, for example, a random forest Classification model including multiple CART (Classification And Regression Tree). The classification model is trained based on expert labeled data, which are training samples for training the classification model. The training samples include: the score of the historical recommended behavior information corresponding to each training sample is determined manually, and the matching degree of the sample combination information in the training sample and the sample user (namely, the sample label).

Wherein the historical recommendation behavior information comprises at least one of: the method comprises the following steps that compliance evaluation information of a sample user for the sample resource, purchasing behavior information of the sample user for the sample resource, and historical operating behavior information of the sample user for the sample resource, for example, browsing behavior information, collecting behavior information, sharing behavior information and the like of the sample user for the sample resource in an application program, the sample user consults behavior information of the sample resource to a second user, and evaluation information of historical recommendation behavior of the sample user for recommending the sample resource to the second user.

Illustratively, referring to table 1, table 1 is a historical recommended behavior information table provided in an embodiment of the present invention.

TABLE 1

Evaluation index	Compliance evaluation	Act of transacting	User service evaluation	APP behaviors	Consultation behavior	…	Sample label	Degree of matching
									Sample 1	5	5	5	5	5	…	Positive sample	1
Sample No. 2	5	0	3	5	1	…	Positive sample	1
									Sample 3	3	0	2	1	1	…	Discarding samples	-
Sample No. 4	4	0	1	0	3	…	Discarding samples	-
									Sample 5	2	0	2	2	2	…	Negative sample	0
…	…	…	…	…	…	…	…	…

Samples 1 to 5 are initial training samples in the present embodiment. The evaluation index is historical recommendation behavior information in the embodiment of the invention. The compliance evaluation is compliance evaluation information of the sample user for the sample resource in the embodiment of the present invention, for example, the compliance evaluation information of the sample user for the sample resource corresponding to the sample 1 is: and if the sample resource is in compliance, the score of the compliance evaluation information of the sample user for the sample resource corresponding to the sample 1 can be determined to be 5. The transaction behavior represents purchase behavior information of the sample user for the sample resource in the embodiment of the present invention, for example, if the sample user purchases the sample resource corresponding to the sample 1, the score of the purchase behavior information of the sample user for the sample resource corresponding to the sample 1 may be determined to be 5, and if the sample user does not purchase the sample resource corresponding to the sample 2, the score of the purchase behavior information of the sample user for the sample resource corresponding to the sample 2 may be determined to be 0. The user service evaluation represents evaluation information of a historical recommendation behavior of the sample user for recommending the sample resource to the second user in the embodiment of the present invention, for example, the evaluation information of the historical recommendation behavior of the sample user for recommending the sample resource corresponding to sample 1 to the second user is: if yes, the rating of the evaluation information of the historical recommendation behavior of the sample user for recommending the sample resource corresponding to the sample 1 by the second user may be determined to be 5. The APP (Application) behavior represents the historical operating behavior information of the sample user for the sample resource in the embodiment of the present invention, for example, if the sample user collects the sample resource corresponding to the sample 1 in the Application program, the score of the historical operating behavior information of the sample user for the sample resource corresponding to the sample 1 may be determined to be 5. The consultation behavior indicates that the sample user consults the second user for the behavior information of the sample resource in the embodiment of the present invention, for example, if the sample user consults the second user for the sample resource corresponding to the sample 1, the score of the behavior information of the sample resource corresponding to the sample user consulting the second user for the sample 1 may be determined to be 5.

For each sample resource, the server may input, to the classification model, one sample combination information, one sample resource demand information, one sample personal basic information, and the score based on the historical recommended behavior information of the sample resource, to obtain a corresponding sample label, for example, if the sample label of the sample 1 is a positive sample, it indicates that the matching degree between the sample user and the sample combination information in the sample 1 is 1. If the sample label of the sample 5 is a negative sample, it indicates that the matching degree between the sample user and the sample combination information in the sample 5 is 0. If the sample label of sample 3 is a discard sample, it indicates that sample 3 is abnormal data.

The server may then determine a number of first initial training samples for which the sample label represents a match with the sample user, and a number of second initial training samples for which the sample label represents a mismatch with the sample user, and calculate a ratio of the number of first initial training samples to the number of second initial training samples. If the calculated ratio is larger than the preset threshold value, the difference between the number of the first initial training samples and the number of the second initial training samples is small, the effect of training the matching degree prediction model of the initial structure is good, the server can directly determine that the first initial training samples are positive samples, and determine that the second initial training samples are negative samples, and the target training samples are obtained.

If the calculated ratio is not larger than the preset threshold, the difference between the number of the first initial training samples and the number of the second initial training samples is larger, the training effect on the matching degree prediction model of the initial structure is poor, the server determines that the first initial training samples are positive samples, and after the second initial training samples are negative samples, a part of samples can be selected from the third initial training samples to be used as positive samples. The third initial training sample is: and the other training samples except the first initial training sample and the second initial training sample, namely the corresponding sample labels are the initial training samples of the discarded samples.

In one implementation, for each first initial training sample, the server may calculate a similarity between the first initial training sample and each third initial training sample. Then, the third initial training sample with the maximum similarity of the first initial training sample is taken as a positive sample, that is, the matching degree between the sample combination information in the third initial training sample and the sample user is determined to be 1, so as to obtain the target training sample.

Referring to fig. 6, fig. 6 is a flowchart of a resource recommendation method in the prior art. In the related art, when recommending a financial product to a user, social data of the user on a social platform for a period of time (e.g., a week) is acquired, and the social data may represent the user's intention. Then, a product list containing the target financial products is determined and recommended based on the recommendation method and the user intention, namely, keywords associated with the products are extracted from the social data of the user, the products associated with the extracted keywords are determined as the target financial products, and the product list containing the target financial products is obtained. Further, the target financial product may be recommended to the user. Therefore, in the related art, only the keywords representing the intention of the user are acquired, and the financial product recommendation is performed based on the acquired keywords, so that the recommendation effectiveness is low, the recommendation mode is single, the recommendation information for performing the financial product recommendation on the user cannot be determined, and the personalized recommendation cannot be performed for the user.

Referring to fig. 7, fig. 7 is a structural diagram of a resource recommendation system according to an embodiment of the present invention. The resource recommendation system comprises: an employee client (the second client in the foregoing embodiment), a user handset (the first client in the foregoing embodiment), a voice gateway, and a customer service system. The customer service system includes: the system comprises a call task management module, a recommendation model (a matching degree prediction model in the previous embodiment), a training module, a call server, an information acquisition module, a sample screening module and a service database.

The customer service system may be the server in the foregoing embodiment, and the modules in the customer service system that implement different functions (i.e., the call task management module, the call server, the information acquisition module, the sample screening module, and the service database) are modules in the server. When the server is a server cluster in the foregoing embodiment, the client system may be a server cluster. The modules (namely the call task management module, the call server, the information acquisition module, the sample screening module and the service database) for realizing different functions in the customer service system are different servers.

The user's handset may send the first user's voice stream (i.e., first audio data) to the call server through the voice gateway. The voice gateway is used for realizing the voice communication between the user mobile phone and the employee client under the condition that the network for the voice communication between the user mobile phone and the employee client is different from the network for the voice communication between the user mobile phone and the employee client. For example, when the user mobile phone performs voice communication according to the mobile phone number, a Network where the user mobile phone performs voice communication is a PSTN (Public Switched Telephone Network) Network. When the employee client performs voice communication based on the application program, the network that performs voice communication is an IP (Internet Protocol) network.

When receiving the first audio data, the call server sends the first audio data to the information acquisition module and sends a user identifier of the first user to the call task management module, and the call task management module generates a task work order corresponding to the first user based on the user identifier of the first user and sends the task work order to the employee client. The task work order is used for displaying information received by the employee client, for example, when communication is established with a mobile phone of the user, first personal basic information of the first user is displayed, and when recommendation is performed on the first user, target recommendation information of a target resource corresponding to the first user is displayed. And the call task management module sends the user identification of the first user to the information acquisition module. The information acquisition module acquires personal basic information corresponding to the user identification of the first user from the business database to obtain first personal basic information, and acquires information characteristic vectors of each to-be-recommended combination information of each first to-be-recommended resource from the business database.

The information acquisition module can also perform voice recognition on the first audio data to obtain first resource demand information of the first user. And then, inputting the first personal basic information, the first resource demand information and the information characteristic vector of each combined information to be recommended into the recommendation model. The recommendation model can output the matching degree of the first user and each piece of combined information to be recommended. The information acquisition module determines target recommendation information of target resources recommended to the first user based on the matching degree of the first user and each piece of combined information to be recommended, and sends the target recommendation information to the employee client. The employee client may display the target recommendation information, acquire second audio data sent by the second user according to the target recommendation information, and send the voice stream (i.e., the second audio data) to the call server. The call server may send the second audio data to the user's handset. The employee mobile phone may play the second audio data to recommend the target resource to the first user according to the target recommendation information.

In addition, the information acquisition module can also record audio data of voice communication between the user mobile phone and the employee client and send the audio data to the sample screening module. The sample screening module can generate a target training sample for training the matching degree prediction model based on audio data of voice communication between a user mobile phone and the staff client, and sends the target training sample to the training module. The training module can train the matching degree prediction model of the initial structure based on the target training sample to obtain a trained matching degree prediction model (namely, a recommendation model).

Based on the processing, the first resource demand information can embody the characteristics of the resource required by the first user, the first personal basic information can embody the real information of the first user, and the to-be-recommended combined information of the first to-be-recommended resource includes the resource characteristics of the first to-be-recommended resource. Correspondingly, the matching degree of the first user and the combined information to be recommended, which is determined based on the first personal basic information, the first resource demand information and the combined information to be recommended, can embody the degree that the first resource to be recommended meets the real demand of the first user. Furthermore, the first alternative resource determined based on the matching degree of the first user and the combined information to be recommended of each first resource to be recommended meets the real requirement of the user, the target resource containing the first alternative resource is recommended to the first user, and the effectiveness of the recommended resource can be improved.

Referring to fig. 8, fig. 8 is a structural diagram of another resource recommendation system according to an embodiment of the present invention. The resource recommendation system comprises: the system comprises a call task management module, a call server, a user mobile phone, a service database, an information acquisition module, a recommendation model and a sample screening module.

The user handset may send first audio data of the first user to the call server. When receiving the first audio data, the call server sends the first audio data to the information acquisition module, and sends the user identifier of the first user to the call task management module, and the call task management module sends the user identifier of the first user to the information acquisition module.

The information acquisition module acquires task information, namely acquires a user identifier of the first user. The information acquisition module carries out voice stream acquisition and carries out voice to text conversion. Then, the information acquisition module collects information and inputs the collected information into the recommendation model, namely the information acquisition module performs voice recognition on the received first audio data to obtain first resource requirement information of the first user, obtains first personal basic information corresponding to the user identification of the first user, and inputs the first personal basic information and the first resource requirement information into the recommendation model. The recommendation model can process the first personal basic information, the first resource demand information and the information characteristic vector of each to-be-recommended combined information of each to-be-recommended resource, and output the matching degree of the first user and each to-be-recommended combined information. And then, the information acquisition module determines the target resource recommended to the user based on the matching degree of the first user and each piece of combined information to be recommended, and sends the target recommendation information of the target resource to the second client. And recommending the target resource to the first user through the second client according to the target recommendation information.

In addition, the information acquisition module is used for acquiring user characteristics and quantifying user behaviors, and the information acquisition module can also be used for acquiring behaviors and processing texts. Then, generating a sample, sending the generated training sample to a sample screening module, namely, an information acquisition module records audio data of voice communication between a user mobile phone of a sample user and an employee client, performing voice recognition on the recorded audio data to obtain sample resource demand information and historical recommendation information of sample resources, generating an initial training sample for training a matching degree prediction model based on sample personal basic information, the sample resource demand information and sample combination information of the sample resources (namely, resource characteristics and the historical recommendation information of the sample resources), and sending the initial training sample to the sample screening module. The sample screening module can determine a target training sample from the initial training sample, and train the matching degree prediction model of the initial structure based on the target training sample to obtain a trained matching degree prediction model (namely, a recommendation model).

Referring to fig. 9, fig. 9 is a flowchart of another resource recommendation method according to an embodiment of the present invention. The resource recommendation method is applied to a resource recommendation system. The resource recommendation system comprises: the system comprises a user client, a staff client, a call server, a task management module, an information acquisition module, a recommendation model, a classification module and a training module. The employee client is the second client in the foregoing embodiment, and the user client is the first client in the foregoing embodiment.

The user client initiates a call to the call server, and the call server establishes a call between the user client and the employee client. That is, the user client sends a voice communication request for the employee client to the call server. And when receiving the voice communication request, the call server establishes communication connection between the user client and the employee client. The call server sends the task information to the task management module, that is, the call server sends the user identifier of the first user to the task management module. The task management module identifies the user and inquires user information from the database, that is, the task management module obtains the personal basic information corresponding to the user identifier of the first user from the service database to obtain the first personal basic information, and sends the user inherent information (namely, the first personal basic information) to the information acquisition module. When receiving the voice stream (namely the first audio data) of the user client, the call server sends the voice stream (namely the first audio data) to the information acquisition module.

The information acquisition module carries out information processing to obtain quantitative characteristics, namely the information acquisition module carries out voice recognition on the first audio data to obtain first resource demand information of the first user. The information acquisition module inputs the inherent characteristics and the intention characteristics of the user into the recommendation model, namely the information acquisition module inputs the first personal basic information and the first resource demand information into the recommendation model. And performing feature splicing processing on the recommendation model to obtain a user vector, calculating the similarity between the user vector and the product vector as recommendation degrees to obtain a sorted list, and returning the first N recommendation degrees which are higher than a threshold value. The recommendation model generates a user feature vector of the first user based on the first personal basic information and the first resource demand information, and calculates similarity between the user feature vector of the first user and information feature vectors of the combined information to be recommended of the first resource to be recommended, so as to obtain matching degrees of the first user and the combined information to be recommended. The information acquisition module determines N target resources recommended to the first user based on the matching degree of the first user and each piece of combined information to be recommended, the matching degree of the target combined information of the target resources and the first user is larger than a matching degree threshold value, and a recommendation list containing the target recommendation information of the target resources is sent to the staff client.

And the employee client and the user client perform product recommendation and feedback, namely, the employee client recommends the target resource to the first user according to the target recommendation information, and after the resource recommendation is completed, the call is ended, namely, the voice communication between the user client and the employee client is stopped.

In addition, the information acquisition module can also acquire user behaviors within a period of time, generate a new training sample based on the acquired user behaviors, and send the new training sample to the classification module. And the classification module classifies the new training sample to obtain a classified sample. That is, the information acquisition module generates an initial training sample based on the audio data of the resource recommendation from the second user to the first user, and determines a target training sample from the initial training sample through the classification module. The training module can train the matching degree prediction model of the initial structure based on the target training sample to obtain a trained matching degree prediction model (namely, a recommendation model) so as to realize the periodic updating of the online model.

Referring to fig. 10, fig. 10 is a schematic diagram of a resource recommendation method according to an embodiment of the present invention. The matching degree prediction model comprises a user characteristic processing module, a resource characteristic processing module and a matching degree calculation module.

In the off-line training phase, the server generates a target training sample containing user inherent features (namely sample personal basic information), intention features (namely sample resource demand information), resource inherent features (namely resource features of sample resources), and introduction features (namely historical recommendation information of the sample resources).

The user characteristic processing module can perform Onehot coding on the non-sequence user inherent characteristics of the sample user and perform Embedding coding on the sequence user inherent characteristics of the sample user to obtain the corresponding characteristic vector. The user characteristic processing module can conduct Embedding coding on the intention characteristics of the sample user to obtain corresponding characteristic vectors. And then, performing feature fusion on the two obtained feature vectors (namely splicing the two obtained feature vectors) to obtain a user feature vector of the sample user. The user feature processing module can also perform dimension reduction on the user feature vectors of the sample users through MLP to obtain the sample user feature vectors of the sample users with the specified length.

The resource feature processing module can perform Onehot coding on the non-sequence resource inherent features of the sample resources and perform Embedding coding on the sequence resource inherent features of the sample resources to obtain corresponding feature vectors. The second fusion processing model can perform Embedding coding on the introducer characteristics of the sample resources to obtain corresponding characteristic vectors. Then, feature fusion is performed on the two obtained feature vectors (that is, the two obtained feature vectors are spliced), so that an information feature vector of one sample combination information of the sample resource is obtained. The resource feature processing module can also perform dimension reduction on the information feature vector of one sample combination information of the sample resource through MLP to obtain a sample information feature vector with a specified length.

Then, the matching degree calculation module calculates the similarity between the sample information feature vector and the sample user feature vector based on the above equation (1) as the matching degree (i.e., the predicted matching degree) of the obtained sample combination information and the sample user. The server may calculate a loss function value representing a difference between the predicted matching degree and a sample matching degree, the sample matching degree being a matching degree represented by a sample label of the target training sample. The server may update parameters based on the calculated loss function value, that is, adjust model parameters of the matching degree prediction model of the initial structure based on the calculated loss function value until a preset convergence condition is reached, so as to obtain a trained matching degree prediction model.

And then, issuing a model to obtain a fusion processing model, namely applying a trained matching degree prediction model to perform online recommendation, wherein in an online recommendation stage, a server determines an intention characteristic (first resource demand information) of a first user based on real-time voice interaction, and inputs the inherent characteristic (first personal basic information) and the intention characteristic of the first user into the fusion processing model. The fusion processing module acquires the information characteristic vector of each to-be-recommended combined information of each first to-be-recommended resource from the resource vector library, and obtains the matching degree of each first user and each to-be-recommended combined information based on the inherent features and the intention features of the first user and the information characteristic vector of each to-be-recommended combined information. Then, the matching degree is sorted to obtain a recommendation list containing products and recommendation words, that is, the server determines, from the first resources to be recommended, target resources of which the matching degrees corresponding to the combined information to be recommended meet preset conditions according to the sequence of the matching degrees of the first users from large to small, and obtains the recommendation list containing the target resources and the target recommendation information.

Referring to fig. 11, fig. 11 is a flowchart of a training sample generation method according to an embodiment of the present invention.

The server generates a training sample based on the user inherent characteristics (i.e. sample personal basic information) of the sample user, the intention characteristics (i.e. sample resource demand information), the resource inherent characteristics (i.e. resource characteristics) of the sample resource, and the recommender characteristics (i.e. historical recommendation information), and generates a label (i.e. sample label) of the training sample based on the evaluation index (i.e. historical recommendation behavior information).

The user-inherent characteristics of the sample user include: basic information (e.g., age of the sample user), income (e.g., annual income information of the sample user), total assets (e.g., deposit amount of the sample user), risk preference (i.e., financial risk level preferred by the sample user), and historical transaction sequence (i.e., information of financial products that the sample user has historically transacted), etc.

And the recording file is sample audio data for resource recommendation of the second user to the sample user, and text conversion is performed on the sample audio data to obtain a dialog text between the second user and the sample user. The server can perform text processing on the obtained conversation text, and obtains the intention characteristics of the sample user by combining the emotion characteristics of the sample user. The server extracts the characteristics of the sample audio data to obtain the emotional characteristics of the sample user, and extracts the text containing the sample resource demand information of the sample user from the dialog text of the second user and the sample user to obtain the intention characteristics of the sample user.

The resource intrinsic characteristics of the sample resource include: basic information (e.g., name, code of sample resource), a sequence of rates of return (e.g., rate of return for a sample resource for one year), a risk level (i.e., financial risk level), and the like. The server can perform text processing on the dialog text of the second user and the sample user to obtain the recommendation language characteristics (namely historical recommendation information) of the sample resources. The server may further perform compliance quality inspection on the dialog texts of the second user and the sample user to obtain compliance evaluation of the sample resource, that is, extract the compliance evaluation of the sample resource from the dialog texts of the second user and the sample user.

The evaluation indexes include: compliance evaluation (i.e., compliance evaluation information of the sample user for the sample resource), transaction behavior (i.e., purchase behavior information of the sample user for the sample resource), operation behavior (i.e., historical operation behavior information of the sample user for the sample resource), consultation behavior (i.e., behavior information of the sample user for consulting the second user about the sample resource), user service evaluation (i.e., evaluation information of historical recommendation behavior of the sample user for recommending the sample resource to the second user), and the like.

Referring to fig. 12, fig. 12 is a flowchart of a training sample generation method according to an embodiment of the present invention.

The sample set S can be expressed as

，

Represents the initial training samples in the sample set S, and a represents the number of initial training samples. The discarded sample set D may be represented as

，

Indicating that the third initial training sample in the sample set D is discarded, and b indicates the number of the third initial training samples. Positive sample set Pmax tableShown as

，

Representing the first initial training sample in the positive sample set P, c represents the number of first initial training samples. The negative sample set N can be expressed as

，

Representing the second initial training sample in the negative sample set N, and d represents the number of the second initial training samples. The sample classification model is the classification model in the previous embodiment.

The server determines sample labels of the initial training samples in the sample set S based on the classification model, and determines a positive sample set P composed of first initial training samples matched with the sample user, a negative sample set N composed of second initial training samples unmatched with the sample user and a discarded sample set D composed of third initial training samples with the sample labels representing discarded samples. In the case that the ratio of the number of the first initial training samples to the number of the second initial samples is not greater than the preset threshold, after the first initial training samples and the second initial training samples are determined as the target training samples, the server supplements the samples based on the discarded sample set D and the positive sample set P, that is, for each first initial training sample, the server calculates the similarity between the first initial training sample and each third initial training sample. Then, the third initial training sample with the maximum similarity to the first initial training sample is determined as a positive sample, and a target training sample is obtained.

Corresponding to the method embodiment of fig. 1, referring to fig. 13, fig. 13 is a structural diagram of a resource recommendation device provided in an embodiment of the present invention, where the device is applied to a server, and the device includes: a first obtaining module 1301, configured to obtain personal basic information of a first user as first personal basic information, and obtain resource requirement information of the first user as first resource requirement information; a second obtaining module 1302, configured to obtain, for each first resource to be recommended, a resource feature of the first resource to be recommended and multiple different pieces of historical recommendation information; the historical recommendation information of the first resource to be recommended is as follows: the method comprises the steps that audio data generated when the first resource to be recommended is recommended to a user in a first historical time period are obtained; the first determining module 1303 is configured to, for each piece of combined information to be recommended of the first resource to be recommended, process the first personal basic information, the first resource demand information, and the combined information to be recommended based on a pre-trained matching degree prediction model to obtain a matching degree between the first user and the combined information to be recommended; the combined information to be recommended comprises the resource characteristics of the first resource to be recommended and historical recommendation information; the historical recommendation information contained in each two pieces of combined information to be recommended is different; the matching degree prediction model is as follows: training based on sample personal basic information of a sample user, sample resource demand information and sample combination information of sample resources to obtain the personal basic information and the sample resource demand information; a second determining module 1304, configured to determine, from the first resources to be recommended, a resource whose matching degree corresponding to the combined information to be recommended meets a preset condition, as a first alternative resource; a recommending module 1305, configured to recommend a target resource to a first user; wherein the target resource comprises a first alternative resource.

Optionally, the first obtaining module 1301 is specifically configured to receive first audio data, which is sent by a first client that a first user logs in, and is used for obtaining a resource; and performing voice recognition on the first audio data to obtain resource demand information of the first user, wherein the resource demand information is used as the first resource demand information.

Optionally, the matching degree prediction model includes: the system comprises a resource feature processing module, a user feature processing module and a matching degree calculating module; the resource characteristic processing module comprises: the system comprises a first feature mapping module, a first feature fusion module and a first multilayer perceptron; the user characteristic processing module comprises: the system comprises a second feature mapping module, a second feature fusion module and a second multilayer perceptron;

the first determining module 1303 is specifically configured to, for each piece of combined information to be recommended of the first resource to be recommended, perform mapping processing on resource features in the combined information to be recommended through the first feature mapping module to obtain a first feature vector, and perform mapping processing on historical recommendation information in the combined information to be recommended to obtain a second feature vector; splicing the first feature vector and the second feature vector through a first feature fusion module to obtain a third feature vector; mapping the third eigenvector according to the specified length through the first multilayer perceptron to obtain the information eigenvector of the combined information to be recommended; mapping the first personal basic information through a second feature mapping module to obtain a fourth feature vector, and mapping the first resource demand information to obtain a fifth feature vector; splicing the fourth feature vector and the fifth feature vector through a second feature fusion module to obtain a sixth feature vector; mapping the sixth feature vector according to the specified length through a second multilayer perceptron to obtain a user feature vector of the first user; and calculating the similarity between the user characteristic vector of the first user and the information characteristic vector of the combined information to be recommended through a matching degree calculation module to serve as the matching degree of the first user and the combined information to be recommended.

Optionally, the recommending module 1305 is specifically configured to send target recommendation information of the target resource to a second client that the second user logs in, so that the second client displays the received target recommendation information, obtains second audio data sent by the second user according to the target recommendation information, and sends the second audio data to the server; the target recommendation information of the target resource comprises: historical recommendation information in the combined information of which the corresponding matching degree of the target resource meets the preset condition; and when the second audio data is received, sending the second audio data to the first client so that the first client plays the second audio data.

Optionally, the apparatus further comprises: the processing module is configured to determine, before the recommending module 1305 performs recommending a target resource to a first user, a plurality of resources without history recommendation information as a second resource to be recommended; for each second resource to be recommended, mapping the resource characteristics of the second resource to be recommended and preset recommendation information to obtain a resource characteristic vector of the second resource to be recommended; calculating the similarity between the resource feature vector of the second resource to be recommended and the information feature vector of the target combination information of each first alternative resource to obtain the similarity between the second resource to be recommended and the first alternative resource; the matching degree corresponding to the target combination information of the first alternative resource meets a preset condition; calculating the mean value of the similarity between the second resource to be recommended and each first alternative resource to obtain the mean value of the similarity corresponding to the second resource to be recommended; determining a first number of resources from the second resources to be recommended as second alternative resources according to the sequence of the corresponding similarity mean values from large to small; wherein the target resource further comprises a second alternative resource.

Optionally, the apparatus further comprises: the training module is used for acquiring sample audio data generated when a second user recommends resources to the sample user in a second historical time period aiming at each sample resource, and performing voice recognition on the sample recommended audio data to obtain historical recommendation information of the sample resource and sample resource demand information of the sample user corresponding to the sample resource; generating an initial training sample based on sample combination information of the sample resource, sample resource demand information and sample personal basic information of a sample user; wherein one sample combination information includes: the resource characteristics of the sample resource and historical recommendation information; the historical recommendation information contained in each two sample combination information is different; determining a positive sample matched with the sample user and a negative sample not matched with the sample user from each initial training sample based on the sample label in each initial training sample to obtain a target training sample; wherein the sample label in an initial training sample represents: matching degree of sample combination information in the initial training sample and a sample user; and adjusting model parameters of the matching degree prediction model of the initial structure based on the target training sample until a preset convergence condition is reached, so as to obtain the trained matching degree prediction model.

Optionally, the training module is further configured to, before obtaining a target training sample by determining, based on the sample label in each initial training sample, a positive sample matched with the sample user and a negative sample unmatched with the sample user from each initial training sample, perform, for each initial training sample, inputting the initial training sample and historical recommended behavior information of the sample resource corresponding to the initial training sample to a pre-trained classification model, and obtaining the sample label of the initial training sample; wherein the historical recommendation behavior information of a sample resource comprises at least one of: the method comprises the following steps that compliance evaluation information of a sample user for the sample resource, purchasing behavior information of the sample user for the sample resource, historical operation behavior information of the sample user for the sample resource, the behavior information of the sample user for consulting the second user about the sample resource, and satisfaction evaluation information of historical recommendation behaviors of the sample user for recommending the sample resource to the second user are obtained;

the training module is specifically used for determining the number of first initial training samples of which the sample labels represent the samples matched with the sample users, and the number of second initial training samples of which the sample labels represent the samples unmatched with the sample users; calculating the ratio of the number of the first initial training samples to the number of the second initial training samples; if the calculated ratio is larger than a preset threshold value, determining that the first initial training sample is a positive sample, and determining that the second initial training sample is a negative sample to obtain a target training sample; if the calculated ratio is not larger than the preset threshold value, calculating the similarity between each first initial training sample and each third initial training sample aiming at each first initial training sample; wherein the third initial training sample is: training samples other than the first initial training sample and the second initial training sample; and determining the first initial training sample and a third initial training sample with the maximum similarity to the first initial training sample as positive samples, and determining the second initial training sample as a negative sample to obtain a target training sample.

Based on the resource recommendation device provided by the embodiment of the invention, the first resource demand information can embody the characteristics of the resource required by the first user, the first personal basic information can embody the real information of the first user, and the to-be-recommended combined information of the first to-be-recommended resource includes the resource characteristics of the first to-be-recommended resource. Correspondingly, the matching degree of the first user and the combined information to be recommended, which is determined based on the first personal basic information, the first resource demand information and the combined information to be recommended, can embody the degree that the first resource to be recommended meets the real demand of the first user. Furthermore, the first alternative resource determined based on the matching degree of the first user and the combined information to be recommended of each first resource to be recommended meets the real requirement of the user, the target resource containing the first alternative resource is recommended to the first user, and the effectiveness of the recommended resource can be improved. In addition, the historical recommendation information contained in each piece of combined information to be recommended of the first resource to be recommended is different, and the matching degree between one piece of combined information to be recommended and the first user can represent the degree that the historical recommendation information in the combined information to be recommended meets the real requirement of the first user. And determining target recommendation information of the target resource according with the real requirement of the first user based on the matching degree of each to-be-recommended combination information and the first user. The target resource is recommended to the first user based on the target recommendation information, personalized resource recommendation can be performed on the basis of the target recommendation information meeting the real requirements of the users aiming at different users, and the effectiveness of the recommended resource is further improved.

An embodiment of the present invention further provides an electronic device, as shown in fig. 14, including a processor 1401, a communication interface 1402, a memory 1403, and a communication bus 1404, where the processor 1401, the communication interface 1402, and the memory 1403 complete communication with each other through the communication bus 1404, and the memory 1403 is used for storing a computer program; the processor 1401 is configured to implement the steps of any of the resource recommendation methods described above when executing the program stored in the memory 1403.

The communication bus mentioned in the electronic device may be a Peripheral Component Interconnect (PCI) bus, an Extended Industry Standard Architecture (EISA) bus, or the like. The communication bus may be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, only one thick line is shown, but this is not intended to represent only one bus or type of bus. The communication interface is used for communication between the electronic equipment and other equipment. The Memory may include a Random Access Memory (RAM) or a Non-Volatile Memory (NVM), such as at least one disk Memory. Alternatively, the memory may be at least one memory device located remotely from the processor. The Processor may be a general-purpose Processor, including a Central Processing Unit (CPU), a Network Processor (NP), and the like; but also 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.

In another embodiment of the present invention, a computer-readable storage medium is further provided, in which a computer program is stored, and the computer program, when executed by a processor, implements the steps of any of the resource recommendation methods described above.

In yet another embodiment, a computer program product containing instructions is provided, which when run on a computer causes the computer to perform any of the resource recommendation methods in the above embodiments.

In the above embodiments, all or part of the implementation may be realized by software, hardware, firmware, or any combination thereof. When implemented in software, may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When loaded and executed on a computer, cause the processes or functions described in accordance with the embodiments of the invention to be performed in whole or in part. The computer may be a general purpose computer, a special purpose computer, a network of computers, or other programmable device. The computer instructions may be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device, such as a server, a data center, etc., that includes one or more available media. The usable medium may be a magnetic medium (e.g., floppy Disk, hard Disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., Solid State Disk (SSD)), among others.

It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in the process, method, article, or apparatus that comprises the element.

All the embodiments in the present specification are described in a related manner, and the same and similar parts among the embodiments may be referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, for the apparatus, system, electronic device, computer-readable storage medium, and computer program product embodiments, the description is relatively simple as it is substantially similar to the method embodiments, and reference may be made to some descriptions of the method embodiments for related points.

The above description is only for the preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention shall fall within the protection scope of the present invention.

Claims

1. A resource recommendation method is applied to a server and comprises the following steps:

acquiring personal basic information of a first user as first personal basic information, and acquiring resource demand information of the first user as first resource demand information;

aiming at each first resource to be recommended, acquiring resource characteristics of the first resource to be recommended and a plurality of different historical recommendation information; the historical recommendation information of the first resource to be recommended is as follows: the method comprises the steps that audio data generated when a first resource to be recommended is recommended to a user in a first historical time period are obtained;

processing the first personal basic information, the first resource demand information and the combined information to be recommended based on a pre-trained matching degree prediction model aiming at each combined information to be recommended of the first resource to be recommended to obtain the matching degree of the first user and the combined information to be recommended; the combined information to be recommended comprises the resource characteristics of the first resource to be recommended and historical recommendation information; the historical recommendation information contained in each two pieces of combined information to be recommended is different; the matching degree prediction model is as follows: the method comprises the steps that training is carried out on the basis of sample personal basic information of sample users, sample resource demand information and sample combination information of sample resources;

determining a resource of which the matching degree corresponding to the combined information to be recommended meets a preset condition from each first resource to be recommended as a first alternative resource;

recommending a target resource to the first user; wherein the target resource comprises the first alternative resource.

2. The method according to claim 1, wherein the acquiring the resource demand information of the first user as the first resource demand information comprises:

receiving first audio data which are sent by a first client logged in by the first user and used for acquiring resources;

and performing voice recognition on the first audio data to obtain resource demand information of the first user, wherein the resource demand information is used as first resource demand information.

3. The method of claim 1, wherein the match prediction model comprises: the system comprises a resource feature processing module, a user feature processing module and a matching degree calculating module; the resource feature processing module comprises: the system comprises a first feature mapping module, a first feature fusion module and a first multilayer perceptron; the user characteristic processing module comprises: the second characteristic mapping module, the second characteristic fusion module and the second multilayer perceptron are arranged in the same plane;

the method for processing the first personal basic information, the first resource demand information and the combined information to be recommended according to each combined information to be recommended of the first resource to be recommended based on a pre-trained matching degree prediction model to obtain the matching degree of the first user and the combined information to be recommended comprises the following steps:

for each piece of combined information to be recommended of the first resource to be recommended, mapping the resource features in the combined information to be recommended through the first feature mapping module to obtain a first feature vector, and mapping historical recommendation information in the combined information to be recommended to obtain a second feature vector;

splicing the first feature vector and the second feature vector through the first feature fusion module to obtain a third feature vector;

mapping the third feature vector according to the specified length through the first multilayer perceptron to obtain an information feature vector of the combined information to be recommended;

mapping the first personal basic information through the second feature mapping module to obtain a fourth feature vector, and mapping the first resource demand information to obtain a fifth feature vector;

splicing the fourth feature vector and the fifth feature vector through the second feature fusion module to obtain a sixth feature vector;

mapping the sixth feature vector according to the specified length through the second multilayer perceptron to obtain the user feature vector of the first user;

and calculating the similarity between the user characteristic vector of the first user and the information characteristic vector of the combined information to be recommended by the matching degree calculation module to serve as the matching degree of the first user and the combined information to be recommended.

4. The method of claim 1, wherein recommending the target resource to the first user comprises:

sending target recommendation information of the target resource to a second client logged in by a second user, so that the second client displays the received target recommendation information, acquires second audio data sent by the second user according to the target recommendation information, and sends the second audio data to the server; wherein the target recommendation information of the target resource comprises: historical recommendation information in the combined information of which the corresponding matching degree of the target resource meets the preset conditions;

and when the second audio data is received, sending the second audio data to the first client so that the first client plays the second audio data.

5. The method of claim 1, wherein prior to the recommending a target resource to the first user, the method further comprises:

determining a plurality of resources without history recommendation information as second resources to be recommended;

for each second resource to be recommended, mapping the resource characteristics of the second resource to be recommended and preset recommendation information to obtain a resource characteristic vector of the second resource to be recommended;

calculating the similarity between the resource feature vector of the second resource to be recommended and the information feature vector of the target combination information of each first alternative resource to obtain the similarity between the second resource to be recommended and the first alternative resource; the matching degree corresponding to the target combination information of the first alternative resource meets the preset condition;

calculating the mean value of the similarity between the second resource to be recommended and each first alternative resource to obtain the mean value of the similarity corresponding to the second resource to be recommended;

determining a first number of resources from the second resources to be recommended as second alternative resources according to the sequence of the corresponding similarity mean values from large to small; wherein the target resource further comprises the second alternative resource.

6. The method according to claim 1, wherein the training process of the matching degree prediction model comprises the following steps:

for each sample resource, obtaining sample audio data generated when the second user recommends resources to the sample user in a second historical time period, and performing voice recognition on the sample audio data to obtain historical recommendation information of the sample resource and sample resource demand information of the sample user corresponding to the sample resource;

generating an initial training sample based on sample combination information of the sample resource, sample resource demand information and sample personal basic information of the sample user; wherein one sample combination information includes: the resource characteristics of the sample resource and historical recommendation information; the historical recommendation information contained in each two sample combination information is different;

determining a positive sample matched with the sample user and a negative sample unmatched with the sample user from each initial training sample based on the sample label in each initial training sample to obtain a target training sample; wherein the sample label in one initial training sample represents: matching degree of sample combination information in the initial training sample and the sample user;

and adjusting model parameters of the matching degree prediction model of the initial structure based on the target training sample until a preset convergence condition is reached, and obtaining the trained matching degree prediction model.

7. The method of claim 6, wherein before the determining, based on the sample label in each initial training sample, a positive sample matching the sample user and a negative sample not matching the sample user from each initial training sample to obtain a target training sample, the method further comprises:

for each initial training sample, inputting the initial training sample and historical recommended behavior information of sample resources corresponding to the initial training sample into a pre-trained classification model to obtain a sample label of the initial training sample; wherein the historical recommended behavior information of a sample resource comprises at least one of: the compliance evaluation information of the sample user for the sample resource, the purchase behavior information of the sample user for the sample resource, and the historical operation behavior information of the sample user for the sample resource, the behavior information of the sample user for consulting the second user about the sample resource, and the satisfaction evaluation information of the historical recommendation behavior of the sample user for recommending the sample resource to the second user;

the determining, based on the sample label in each initial training sample, a positive sample matched with the sample user and a negative sample not matched with the sample user from each initial training sample to obtain a target training sample includes:

determining a number of first initial training samples for which a sample label representation matches the sample user, and a number of second initial training samples for which a sample label representation does not match the sample user;

calculating a ratio of the number of the first initial training samples to the number of the second initial training samples;

if the calculated ratio is larger than a preset threshold value, determining that the first initial training sample is a positive sample, and determining that the second initial training sample is a negative sample to obtain a target training sample;

if the calculated ratio is not larger than the preset threshold value, calculating the similarity between each first initial training sample and each third initial training sample aiming at each first initial training sample; wherein the third initial training sample is: training samples other than the first initial training sample and the second initial training sample; and determining the first initial training sample and a third initial training sample with the maximum similarity to the first initial training sample as positive samples, and determining the second initial training sample as a negative sample to obtain a target training sample.

8. A resource recommendation device, characterized in that the device is applied to a server, and the device comprises:

the first acquisition module is used for acquiring personal basic information of a first user as first personal basic information and acquiring resource demand information of the first user as first resource demand information;

the second obtaining module is used for obtaining the resource characteristics of each first resource to be recommended and a plurality of different historical recommendation information; the historical recommendation information of the first resource to be recommended is as follows: the method comprises the steps that audio data generated when a first resource to be recommended is recommended to a user in a first historical time period are obtained;

the first determining module is used for processing the first personal basic information, the first resource demand information and the combined information to be recommended according to each combined information to be recommended of the first resource to be recommended based on a pre-trained matching degree prediction model to obtain the matching degree of the first user and the combined information to be recommended; the combined information to be recommended comprises the resource characteristics of the first resource to be recommended and historical recommendation information; the historical recommendation information contained in each two pieces of combined information to be recommended is different; the matching degree prediction model is as follows: training based on sample personal basic information of a sample user, sample resource demand information and sample combination information of sample resources to obtain the personal basic information and the sample resource demand information;

the second determining module is used for determining the resource of which the matching degree corresponding to the combined information to be recommended meets the preset condition from each first resource to be recommended as a first alternative resource;

a recommending module for recommending a target resource to the first user; wherein the target resource comprises the first alternative resource.

9. A resource recommendation system, the system comprising: first client, second client and server, wherein:

the server is used for performing voice recognition on the first audio data when the first audio data are received to obtain first resource demand information of the first user; aiming at each first resource to be recommended, acquiring resource characteristics of the first resource to be recommended and a plurality of different historical recommendation information; wherein, a history recommendation information of the first resource to be recommended is: the method comprises the steps that audio data generated when the first resource to be recommended is recommended to a user in a first historical time period are obtained; processing the first personal basic information of the first user, the first resource demand information and the combined information to be recommended based on a pre-trained matching degree prediction model aiming at each combined information to be recommended of the first resource to be recommended to obtain the matching degree of the first user and the combined information to be recommended; the combined information to be recommended comprises the resource characteristics of the first resource to be recommended and historical recommendation information; the historical recommendation information contained in each two pieces of combined information to be recommended is different; the matching degree prediction model is as follows: the method comprises the steps that training is carried out on the basis of sample personal basic information of sample users, sample resource demand information and sample combination information of sample resources; determining a resource with matching degree meeting a preset condition corresponding to the combined information to be recommended from each first resource to be recommended as a first alternative resource; target recommendation information of the target resource is sent to the second client logged in by the second user; wherein the target resource comprises the first alternative resource; the target recommendation information of the target resource comprises: historical recommendation information in the combined information of which the corresponding matching degree of the target resource meets the preset conditions;

the first client is further configured to play the second audio data.

10. The electronic equipment is characterized by comprising a processor, a communication interface, a memory and a communication bus, wherein the processor and the communication interface are used for realizing the communication between the processor and the memory through the communication bus;

a memory for storing a computer program;

a processor for implementing the method steps of any of claims 1 to 7 when executing a program stored in the memory.