CN116976988A

CN116976988A - Training method of recommendation model, message recommendation method, device and electronic equipment

Info

Publication number: CN116976988A
Application number: CN202310084751.5A
Authority: CN
Inventors: 张弘; 黄东波
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2023-01-13
Filing date: 2023-01-13
Publication date: 2023-10-31

Abstract

The embodiment of the application provides a training method of a recommendation model, a message recommendation method, a message recommendation device, electronic equipment and a computer readable storage medium, and relates to the field of artificial intelligence. The method comprises the following steps: obtaining message characteristics, object characteristics, target recommendation indexes and first recommendation values; inputting the message characteristics and the object characteristics into a neural network model to obtain a second recommended value, a second sorting result and a first sorting result which are determined by the neural network model; obtaining an actual recommendation index of each sample recommendation message, performing iterative training on the neural network model according to a first difference between a target recommendation index and the actual recommendation index corresponding to each sample recommendation message and a second difference between a second ordering result and a first ordering result of each sample recommendation object, and taking the neural network model after iteration is stopped as a message recommendation model. The embodiment of the application balances the benefits of the recommendation body and the recommendation intermediaries and carries out multi-objective recommendation.

Description

Training method of recommendation model, message recommendation method, device and electronic equipment

Technical Field

The application relates to the technical field of artificial intelligence, in particular to a training method of a recommendation model, a message recommendation method, a message recommendation device, electronic equipment, a computer readable storage medium and a computer program product.

Background

With the rapid development of the Internet, the method brings many convenience to the life of people and simultaneously brings massive data information. In a recommendation scenario, the recommendation system may provide sample recommendation messages of interest to sample recommendation objects based on the message recommendation model, and pertinently recommend the sample recommendation objects to the sample recommendation objects of possible interest.

In a recommendation scene, particularly in an advertisement delivery scene, two types of advertisements (sample recommendation messages), namely contract advertisements and bid advertisements, wherein the contract advertisements are advertisements which are sequenced and recommended based on the recommendation times, the bid advertisements are advertisements which are sequenced and recommended based on the recommendation values, and the research of the contract advertisement delivery algorithm in the industry is mainly an optimization solving problem of distribution under the requirement constraint and the supply constraint. However, in practical application, since recommendation of contract advertisements has the participation of bid advertisements, the current algorithm does not achieve the effect of maximizing the profit of the advertisement platform.

Disclosure of Invention

Embodiments of the present application provide a training method for recommendation model, a message recommendation method, a message recommendation device, an electronic device, a computer readable storage medium and a computer program product, which can solve the above-mentioned problems in the prior art. The technical proposal is as follows:

According to a first aspect of an embodiment of the present application, there is provided a training method of a message recommendation model, including:

obtaining message characteristics of a plurality of sample recommendation messages, object characteristics of a plurality of sample recommendation objects, target recommendation indexes of the sample recommendation messages indicated to a recommendation intermediary by recommendation subjects of each sample recommendation message, and first recommendation values of each sample recommendation object for the sample recommendation messages;

inputting the message characteristics of the plurality of sample recommendation messages and the object characteristics of the plurality of sample recommendation objects into a neural network model, and obtaining a second recommendation value of each sample recommendation message for each sample recommendation object determined by the neural network model and a first sorting result of each sample recommendation message recommended to the sample recommendation object determined based on the second recommendation value;

for each sample recommended object, determining a target sample recommended message recommended to the sample recommended object based on a first sorting result of the sample recommended object, and obtaining an actual recommended index of each sample recommended message according to each sample recommended message serving as a recommended result of the target sample recommended message;

Determining a second sorting result of each sample recommendation message recommended to each sample recommendation object according to the first recommendation value, determining a first difference between a target recommendation index and an actual recommendation index corresponding to each sample recommendation message and a second difference between the second sorting result and the first sorting result of each sample recommendation object, performing iterative training on the neural network model according to the first difference and the second difference until an iteration stopping condition is reached, and taking the neural network model after the iteration stopping as a message recommendation model.

According to a second aspect of the embodiment of the present application, there is provided a message recommendation method, including:

obtaining message characteristics of a plurality of recommended messages and object characteristics of at least one recommended object;

inputting the message characteristics of the plurality of recommended messages and the object characteristics of the plurality of recommended objects into a message recommendation model, and obtaining a first sequencing result of each recommended message recommended to each recommended object, which is determined by the message recommendation model, for each recommended object;

for each recommended object, determining a target recommended message recommended to the recommended object based on a first sorting result of the recommended object, and recommending the target recommended message to the recommended object;

Wherein the message recommendation model is trained based on the method of the first aspect.

According to a third aspect of the embodiment of the present application, there is provided a training apparatus for a message recommendation model, the apparatus including:

the sample data acquisition module is used for acquiring message characteristics of a plurality of sample recommendation messages, object characteristics of a plurality of sample recommendation objects, target recommendation indexes of the sample recommendation messages, indicated to a recommendation intermediary by recommendation subjects of each sample recommendation message, and first recommendation values of each sample recommendation object for the sample recommendation messages;

the sorting module is used for inputting the message characteristics of the plurality of sample recommendation messages and the object characteristics of the plurality of sample recommendation objects into the neural network model, and obtaining a second recommendation value of the sample recommendation object for each sample recommendation message determined by the neural network model and a first sorting result of each sample recommendation message recommended to the sample recommendation object determined based on the second recommendation value, for each sample recommendation object determined by the neural network model;

the index determining module is used for determining target sample recommendation information recommended to each sample recommendation object based on a first sorting result of the sample recommendation object, and obtaining an actual recommendation index of each sample recommendation information according to the recommendation result of each sample recommendation information serving as the target sample recommendation information;

The iterative training module is used for determining a second sorting result of each sample recommendation message recommended to each sample recommendation object according to the first recommendation value, determining a first difference between a target recommendation index and an actual recommendation index corresponding to each sample recommendation message and a second difference between a second sorting result and the first sorting result of each sample recommendation object, performing iterative training on the neural network model according to the first difference and the second difference until an iterative stopping condition is reached, and taking the neural network model after iterative stopping as a message recommendation model.

As an alternative embodiment, the plurality of sample recommendation messages include at least one first-class sample recommendation message and at least one second-class sample recommendation message, wherein the first-class sample recommendation message is a sample recommendation message which is sequenced and recommended according to the recommendation times, and the second-class sample recommendation message is a sample recommendation message which is sequenced and recommended according to the recommendation values;

the target recommendation indexes of the first class sample recommendation message comprise target recommendation times and target effect recommendation times reaching a first preset recommendation effect, and the actual recommendation indexes comprise actual recommendation times and actual effect recommendation times reaching the first preset recommendation effect;

The target recommendation indexes of the second-class sample recommendation message comprise target recommendation cost and target effect recommendation cost for achieving a second preset recommendation effect, and the actual recommendation indexes comprise actual recommendation cost and actual effect recommendation cost for achieving the second preset recommendation effect.

As an alternative embodiment, the index determination module is configured to:

for each first-class sample recommendation message, taking the first-class sample recommendation message as the number of target sample recommendation messages and taking the number of target sample recommendation messages as the actual recommendation times;

and for each first-class sample recommendation message, obtaining an effect coefficient of the first-class sample recommendation message on the corresponding sample recommendation object when the first-class sample recommendation message is used as a target sample recommendation message, wherein the effect coefficient is used for representing the probability that the sample recommendation message is recommended to the sample object to reach a preset recommendation effect, and obtaining actual effect recommendation times according to the effect coefficient of the first-class sample recommendation message on the corresponding sample recommendation object when the first-class sample recommendation message is used as the target sample recommendation message.

As an alternative embodiment, the index determination module is configured to:

for each second-class sample recommendation message, taking the second-class sample recommendation message as a sample recommendation object of the target sample recommendation message as a target sample recommendation object;

Taking a second recommendation value immediately after the second recommendation value of the second class sample recommendation message in the first sequencing result of the target sample recommendation object as a reference value of the target sample recommendation object;

and obtaining the actual recommendation cost of the second-class sample recommendation message according to the reference values of all target sample recommendation objects of the second-class sample recommendation message.

As an alternative embodiment, the index determination module is configured to:

for each second-class sample recommendation message, according to the effect coefficient of the corresponding sample recommendation object when the second-class sample recommendation message is taken as the target sample recommendation message, obtaining the actual effect recommendation times of the second-class sample recommendation message reaching a second preset recommendation effect;

and obtaining the actual effect recommendation cost of the second-class sample recommendation message reaching a second preset recommendation effect according to the actual effect recommendation times and the actual recommendation cost of the second-class sample recommendation message.

As an alternative embodiment, the neural network model includes a first sub-neural network model and a second sub-neural network model;

the first sub-neural network model is used for obtaining a second recommendation value of each sample recommendation object according to the message characteristics of the plurality of sample recommendation messages and the object characteristics of the plurality of sample recommendation objects;

The second sub-neural network model is used for acquiring a first sequencing result of each sample recommendation message recommended to each sample recommendation object according to the second recommendation value of the sample recommendation object for each sample recommendation message according to each sample recommendation object;

the iterative training module is specifically used for:

obtaining a first loss value of a first loss function according to a first difference between a target recommendation index and an actual recommendation index corresponding to each sample recommendation message, and obtaining a second loss value of a second loss function according to a second sorting result of each sample recommendation object and a second difference between the first sorting result;

and updating the model parameters of the first sub-neural network model according to the first loss value and the second loss value, and updating the model parameters of the second sub-neural network model according to the second loss value.

As an alternative embodiment, the first sub-neural network model includes a recommended parameters layer and a value layer;

the recommendation parameter layer is used for obtaining initial recommendation values of the sample recommendation messages according to the message characteristics of the sample recommendation messages and obtaining effect values of preset popularization effects when the sample recommendation messages are recommended according to the message characteristics of the sample recommendation messages and object characteristics of a plurality of sample recommendation objects;

The value layer is used for obtaining a recommendation coefficient of each sample recommendation message for each sample recommendation object, and obtaining a second recommendation value of each sample recommendation message for each sample recommendation object according to the initial recommendation value of the sample recommendation message, the effect value of the sample recommendation message for generating a preset popularization effect when the sample recommendation message is recommended and the recommendation coefficient of the sample recommendation message for the sample recommendation object.

According to a fourth aspect of an embodiment of the present application, there is provided a message recommendation apparatus, including:

the data acquisition module to be processed is used for acquiring message characteristics of a plurality of recommended messages and object characteristics of at least one recommended object;

the ordering determining module is used for inputting the message characteristics of the plurality of recommended messages and the object characteristics of the plurality of recommended objects into the message recommending model, and obtaining a first ordering result of each recommended message recommended to each recommended object, which is determined by the message recommending model, for each recommended object;

the recommendation module is used for determining a target recommendation message recommended to each recommended object based on a first sequencing result of the recommended object, and recommending the target recommendation message to the recommended object;

Wherein the message recommendation model is trained based on the apparatus provided in the third aspect.

According to a fifth aspect of embodiments of the present application there is provided an electronic device comprising a memory, a processor and a computer program stored on the memory, the processor executing the computer program to carry out the steps of the method of the first or second aspect.

According to a sixth aspect of embodiments of the present application there is provided a computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of the method of the first or second aspect.

According to an aspect of an embodiment of the present application, there is provided a computer program product comprising a computer program which, when executed by a processor, implements the steps of the method of the first or second aspect.

The technical scheme provided by the embodiment of the application has the beneficial effects that:

the method comprises the steps of carrying out iterative training on a neural network model to consider differences of two dimensions, wherein one is based on a first difference between a target recommendation index and an actual recommendation index corresponding to each sample recommendation message, training the neural network model by utilizing the first difference, enabling recommendation results determined on the basis of the neural network model to meet the requirements of recommendation subjects as much as possible, balancing benefits of the recommendation subjects and recommendation intermediaries, and the other is a second difference between two sorting results, wherein the second sorting results are obtained on the basis of first recommendation values evaluated by all recommendation subjects, the first sorting results are obtained on the basis of second recommendation values evaluated by the model according to message characteristics and object characteristics, and because the output of the neural network model is the sorting results, on one hand, the sorting results can directly determine whether recommendation messages can be recommended (reflect recommendation times), and on the other hand, the sorting results are derived from recommendation values (reflect recommendation values), so that the scheme skillfully solves the defects that the existing recommendation model only takes one dimension, namely recommendation probability or recommendation value, is not suitable for mixed-ranking scenes.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings that are required to be used in the description of the embodiments of the present application will be briefly described below.

FIG. 1 is a schematic diagram of a bipartite graph according to an embodiment of the present application;

FIG. 2 is a schematic diagram of an application environment of a training method of a message recommendation model according to an embodiment of the present application;

FIG. 3 is a flowchart of a training method of a message recommendation model according to an embodiment of the present application;

FIG. 4 is a schematic diagram of a training process of a message recommendation model according to an embodiment of the present application;

FIG. 5 is a schematic diagram of a training process of another message recommendation model according to an embodiment of the present application;

fig. 6 is a flow chart of a message recommending method according to an embodiment of the present application;

FIG. 7 is a system architecture diagram of a message recommendation system according to an embodiment of the present application;

fig. 8 is a schematic structural diagram of a training device of a message recommendation model according to an embodiment of the present application;

fig. 9 is a schematic structural diagram of a message recommending apparatus according to an embodiment of the present application;

fig. 10 is a schematic structural diagram of an electronic device according to an embodiment of the present application.

Detailed Description

Embodiments of the present application are described below with reference to the drawings in the present application. It should be understood that the embodiments described below with reference to the drawings are exemplary descriptions for explaining the technical solutions of the embodiments of the present application, and the technical solutions of the embodiments of the present application are not limited.

As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless expressly stated otherwise, as understood by those skilled in the art. It will be further understood that the terms "comprises" and "comprising," when used in this specification, specify the presence of stated features, information, data, steps, operations, elements, and/or components, but do not preclude the presence or addition of other features, information, data, steps, operations, elements, components, and/or groups thereof, all of which may be included in the present specification. It will be understood that when an element is referred to as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements may be present. Further, "connected" or "coupled" as used herein may include wirelessly connected or wirelessly coupled. The term "and/or" as used herein indicates that at least one of the items defined by the term, e.g., "a and/or B" may be implemented as "a", or as "B", or as "a and B".

For the purpose of making the objects, technical solutions and advantages of the present application more apparent, the embodiments of the present application will be described in further detail with reference to the accompanying drawings.

First, several terms related to the present application are described and explained:

recommendation message: may be any type of message that needs to be recommended, such as advertisements, news, multimedia content (e.g., public number articles, short videos, long videos), etc. The technical scheme of the embodiment of the application is mainly described by advertisements in the following step.

Recommendation subject: a person, business or entity for providing the recommendation message.

Recommended object: a person, business, or entity for "consuming" the recommended message.

And the recommendation intermediary is used for receiving the recommendation message provided by the recommendation subject and exposing the recommendation message to the recommendation object, and the recommendation intermediary can obtain benefits from the recommendation subject according to the recommendation result.

An advertiser: is the sponsor of the advertisement campaign, is the merchant selling or advertising own products and services on the internet, and is the provider of the alliance marketing advertisement. Any merchant that promotes, sells its products or services may act as an advertiser. The embodiment of the application is also called as a recommended subject.

An advertisement delivery platform: meets the requirement of an advertiser for advertising and provides a platform for flow distribution. The embodiments of the present application are also referred to as recommendation intermediaries.

Contract advertisement: the advertiser and the advertisement delivery platform sign contracts, the advertisement delivery platform is required to play advertisements with preset play amounts to the advertiser in preset types of exposure (also called recommended objects) in preset time, if the contracts are met, the advertiser needs to pay corresponding advertisement delivery fees to the advertisement delivery platform, if the contracts are not met, namely the actual play amounts of the advertisements do not reach the preset play amounts corresponding to the contracts, the advertisement delivery platform needs to pay a certain fee to the advertiser, and when the contract advertisements are played, if the actual play amounts of the advertisements exceed the preset play amounts corresponding to the advertisement delivery amounts, the advertisement delivery platform does not charge additional fees. This means for the advertising system that when an exposure opportunity arrives, there may be multiple advertisers' orders meeting the exposure requirements, the advertising system needs to decide on the contract order that this exposure exposes, and ensure that all other contracts are also completed.

Bid advertisement: an advertisement format paid according to advertisement effect (such as click-through rate, conversion rate, etc.); the advertiser may place a bid (recommended value) for its placed advertisement, and when an exposure request arrives, each bid advertisement whose targeting condition matches the exposure request may compete for the exposure request based on the advertiser's preset bid.

Mixing and sorting: the contract advertisements and the bid advertisements are uniformly ordered and distributed.

The optimization method comprises the following steps: refers to a method for solving the optimization problem. The optimization problem is to determine what values of some selectable variables should take under certain constraint conditions so as to optimize the selected objective function.

Cloud technology (Cloud technology) refers to a hosting technology for integrating hardware, software, network and other series resources in a wide area network or a local area network to realize calculation, storage, processing and sharing of data.

Cloud technology (Cloud technology) is based on the general terms of network technology, information technology, integration technology, management platform technology, application technology and the like applied by Cloud computing business models, and can form a resource pool, so that the Cloud computing business model is flexible and convenient as required. Cloud computing technology will become an important support. Background services of technical networking systems require a large amount of computing, storage resources, such as video websites, picture-like websites, and more portals. Along with the high development and application of the internet industry, each article possibly has an own identification mark in the future, the identification mark needs to be transmitted to a background system for logic processing, data with different levels can be processed separately, and various industry data needs strong system rear shield support and can be realized only through cloud computing.

Big data (Big data) refers to a data set which cannot be captured, managed and processed by a conventional software tool within a certain time range, and is a massive, high-growth-rate and diversified information asset which needs a new processing mode to have stronger decision-making ability, insight discovery ability and flow optimization ability. With the advent of the cloud age, big data has attracted more and more attention, and special techniques are required for big data to effectively process a large amount of data within a tolerant elapsed time. Technologies applicable to big data include massively parallel processing databases, data mining, distributed file systems, distributed databases, cloud computing platforms, the internet, and scalable storage systems.

The contract advertisement is a main stream form of recommended advertisement, namely, the recommended intermediary needs to complete corresponding contract delivery according to the targeting condition of the recommended body in a specified time. In general, this process can be regarded as an allocation problem on the two-part graph g= (I u J, E), as shown in fig. 1. I is a Supply set representing recommended objects under various orientation conditions, and can reach hundreds of millions or billions of dimensions in general, each Supply node I corresponds to different s _i Indicating the total number of exposures under this orientation condition. J is a Demand set, represents contract advertisements of a recommendation body, can reach thousands or tens of thousands of dimensions generally, and each Demand node J corresponds to different d _j Indicating the number of recommendations required for the contract advertisement. And a connection line (i, j) E between the Supply node and the Demand node indicates that the Supply node meets the orientation condition of the Demand node.

CPA (Cost Per Action), per action cost, actions may be registration, interaction, download, order, purchase, etc.; CPA = total cost/conversion, for example, the cost of a recommended message for a recommended person to put into a product over a period of time is $ 6000, the recommended message has an exposure of 600000, a click of 60000, and a conversion (e.g., activation) number of 1200, then the cost per action of the recommended message is: cpa=6000/1200=5 dollars. The total cost is also called the resource occupation amount in the embodiment of the application, and the resource can be money, manpower resource, time and the like.

tCPA, target per action cost, which is set by the recommended topic.

CTR (Click-Through-Rate), which may also be referred to simply as Click Rate, is the actual number of clicks of a recommended message (divided by the display volume (Show content) of the recommended message;

CVR (Conversion Rate), conversion rate of browser click recommendation message to become a valid activation or registration or even payor; CVR = conversion/click-through, an indicator of CPA recommended message effect;

the goal of the allocation problem is to find a viable allocation scheme with a value on each side that indicates that the i-node traffic has x _ij The proportion is assigned to contract advertisement j. How to solve xij is a key issue for contract ad distribution algorithms. To find the optimal x _ij It can be modeled as an optimization problem.

In the selection of objective functions, the related art considers both maximizing advertisement value and minimizing the shortage: the former is to improve the advertising value and the satisfaction of advertisers, and the latter is to ensure the income of the platform. In general, an ideal dispensing target can be set first, for example, as a ratio of demand to available supply-this means uniform exposure to the targeted audience, and then minimizing x _ij Distance to the dispensing target.

It should be noted that one is entirely defined by x _ij The composed solution requires storage space of O (|E|) size. Such a solution is difficult to implement in view of the Supply nodes up to billions of dimensions. The related art demonstrates that when the objective function and constraint are convex, a compact allocation scheme can be obtained, requiring only O (|J|) size space.

Without frequency control, the specific form of the optimization problem is:

wherein: Γ (i) and Γ (j) represent the neighbor node sets of i and j, respectively, θ _ij ＝d _j /S _j To distribute the target S _j ＝∑ _i∈Γ(j) s _i Representation d _j All available flow, V _j Represents the importance of order j, p _j And the penalty coefficient when the order j is in the shortage. The first inequality constraint is called the demand constraint, the second constraint is the supply constraint, and the third and fourth constraints are the non-negative constraints.

The above-mentioned optimization model is specifically solved into two stages, offline and online. Calculating a compact allocation scheme in an off-line stage, wherein the compact allocation scheme consists of dual variables of Demand; on-line stage x is calculated according to dual variables of Demand _ij And then selects an appropriate order for display accordingly.

The existing contract advertisement allocation scheme does not consider the participation of bid advertisements, and cannot enable the allocation scheme to meet the goal of maximizing the media benefit.

In addition, the existing method obtains the playing probability of the advertisement on the exposure request, the bid advertisement gives the bid of the advertisement on the exposure request, the dimension of the bid advertisement and the bid of the advertisement are different, the bid cannot be directly compared, and the distribution scheme of the existing contract advertisement cannot calculate the proper contract order bid.

The application provides a training method of a recommendation model, a message recommendation method, a message recommendation device, electronic equipment, a computer readable storage medium and a computer program product, and aims to solve the technical problems in the prior art.

The technical solutions of the embodiments of the present application and technical effects produced by the technical solutions of the present application are described below by describing several exemplary embodiments. It should be noted that the following embodiments may be referred to, or combined with each other, and the description will not be repeated for the same terms, similar features, similar implementation steps, and the like in different embodiments.

The training method of the message recommendation model provided by the embodiment of the application can be applied to an application environment shown in fig. 2. The terminal 110 communicates with the streaming server 120 through a network, and when the terminal 110 accesses the streaming server 120, the streaming server 120 provides a streaming service to the terminal 110, and it may be understood that the streaming service may be to show a web page to the terminal 110. The streaming media server sends the object characteristics of the terminal 110 to the recommendation server 130, the recommendation server 130 is used for training the obtained message pushing model according to the embodiment of the application, obtaining the pushing message pushed to the terminal 110, and the streaming media server 120 returns the webpage with the recommendation message to the terminal 110.

The server 120 and the recommendation server 130 in the embodiment of the present application may be independent physical servers, or may be a server cluster or a distributed system formed by a plurality of physical servers, or may be a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content distribution networks (Content Delivery Network, CDN), and basic cloud computing services such as big data and an artificial intelligence platform. The terminal may be, but is not limited to, a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc. The terminal and the server may be directly or indirectly connected through wired or wireless communication, and the present application is not limited herein.

The terminal 110 in the embodiment of the present application may be one or more of a smart phone, a vehicle-mounted terminal, a tablet computer, a portable computer, and a desktop computer, which is not limited herein.

The embodiment of the application provides a training method of a message recommendation model, as shown in fig. 3, comprising the following steps:

s101, obtaining message characteristics of a plurality of sample recommendation messages, object characteristics of a plurality of sample recommendation objects, target recommendation indexes of the sample recommendation messages, indicated to a recommendation intermediary by recommendation subjects of each sample recommendation message, and first recommendation values of each sample recommendation object for the sample recommendation messages.

The message characteristics of the sample recommendation message in the embodiment of the application are used for reflecting the relevant characteristics of the sample recommendation message, and can include, for example, the size of the sample recommendation message, the message type (text, audio, picture, dynamic picture, video, etc.), the recommended content type (automobile, cosmesis, movie, entertainment waiting) orientation condition, and the like, and the object characteristics of the sample recommendation object are also used for reflecting the relevant characteristics of the sample recommendation object, and can include, for example, age, gender, location, education level, and the like. The embodiment of the application is not particularly limited to the type of the message feature and the object feature.

In one embodiment, a two-part graph g= (I u J, E) can be generated by obtaining a plurality of sample recommendation messages and matching a plurality of sample recommendation objects through a predetermined orientation relationship, and then the characteristics of the sample recommendation messages and the characteristics of the sample recommendation objects are respectively subjected to characteristic expansion by using the two-part graph, so that the message characteristics of each sample recommendation message and the object characteristics of each sample recommendation object are enriched.

It will be appreciated that the recommendation intermediaries do not randomly proceed in order to recommend the sample recommendation message, they need to allocate according to the target recommendation index indicated by the recommendation entity, and in general, the recommendation intermediaries need to compensate the recommendation entity without reaching the target recommendation index, and how the recommendation entity exceeds the target recommendation index, the recommendation entity does not additionally compensate the recommendation intermediaries, and the recommendation entity needs to instruct the recommendation intermediaries about the recommendation value of the sample recommendation message provided by the recommendation entity to the sample recommendation object, that is, the first recommendation value, provided by the recommendation entity. It can be appreciated that in the field of advertisement delivery, the delivery rule of bid advertisements is to recommend according to the first recommendation value—the higher the first recommendation value, the easier the bid advertisement is delivered.

S102, inputting message characteristics of a plurality of sample recommendation messages and object characteristics of a plurality of sample recommendation objects into a neural network model, and obtaining a second recommendation value of each sample recommendation message and a first sequencing result of each sample recommendation message recommended to the sample recommendation object, which are determined by the neural network model, for each sample recommendation object.

According to the embodiment of the application, the message characteristics of the plurality of sample recommendation messages and the object characteristics of the plurality of sample recommendation objects are used as training samples to be input into the neural network model, and the neural network model is applied to recommendation intermediaries, so that the neural network model is naturally used for recommending the profit of intermediaries, and takes account of the profit of recommendation subjects, namely, the profit of recommendation intermediaries and recommendation subjects can be maximized as much as possible, and the neural network model can obtain the recommendation value of each sample recommendation object evaluated by the recommendation intermediaries for each sample recommendation message, namely, the second recommendation value according to the message characteristics and the object characteristics. The second recommendation value is a recommendation value predicted by the neural network model through machine learning according to the message characteristics and the object characteristics, and the second recommendation value gradually approaches the first recommendation value through training.

The neural network model determines a first ranking result of each sample recommendation message recommended to the sample recommendation object based on a second recommendation value of the sample recommendation object for each sample recommendation message, the first ranking result being based on the first recommendation value, and the first recommendation value being obtained from a message characteristic of the sample recommendation message and an object characteristic of the sample recommendation object. On the other hand, the embodiment of the application also determines the second sorting result of each sample recommendation message recommended to the sample recommendation object according to the first recommendation value of the sample recommendation object for each sample recommendation message. It will be appreciated that whichever ordering model the top-ordered sample recommendation message means that there is a greater probability of being recommended to the sample recommendation object than the bottom-ordered sample recommendation message.

S103, for each sample recommended object, determining a target sample recommended message recommended to the sample recommended object based on a first sorting result of the sample recommended object, and obtaining an actual recommended index of each sample recommended message according to the recommended result of each sample recommended message serving as the target sample recommended message.

When determining a target sample recommendation message recommended to a sample recommendation object according to a first sequencing result, the embodiment of the application has the following constraint:

1. Advertisement comparison constraint for constraining a second recommendation value b of a sample recommendation object i for a sample recommendation message j _ij Second recommendation value b for another sample recommendation message k with sample recommendation object i _ik Is the relation of:

wherein x is _ij Indicating whether the sample recommendation message j is a target sample recommendation message for sample recommendation object i, if so, x _ij 1, if not, x _ij Is 0, gamma _ijk Representing the relative price of sample recommendation message j to sample recommendation message k, ζ, for sample recommendation object i _ijk The relative premium of the sample recommendation message j with respect to the non-target sample recommendation message k for the sample recommendation object i is represented.

It should be noted that, the relative discount price may be obtained based on the second recommendation value of the two sample recommendation messages and the preset discount price coefficient, and the relative overflow price may be obtained based on the second recommendation value of the two sample recommendation messages and the preset overflow price coefficient。A _GD Representing a set of first class sample recommendation messages, A _RTB Representing a set of second class sample recommendation messages.

2. A divalent constraint for constraining a relationship of a difference value of a sample recommendation message j of a highest second recommendation value (i.e., a target sample recommendation message) to a sample recommendation message k of a second highest second recommendation value for a sample recommendation object i:

Wherein delta _ij Representing the difference in the second recommendation value of sample recommendation message j and sample recommendation message k.

3. Assigning variable constraints: that is, for each sample recommendation message, the probability that the sample recommendation message is recommended to one sample recommendation object is 100% at maximum and 0 at minimum, and in the embodiment of the present application, the probability is either 100% or 0.

4. Other non-0 constraints:

the first sequencing result of one sample recommended object is determined by utilizing the second recommendation value determined by the neural network, and when the target sample recommended information of each sample recommended object is determined, the first sequencing result of the sample recommended object is based on the determination, so that the determination of the target sample recommended information of one sample recommended object is related to the recommendation value, namely, the target sample recommended information is suitable for recommending bid advertisements, and meanwhile, the determination of the target sample recommended information is also related to the sequencing result and also directly reacts to the recommendation times of advertisements, so that the target sample recommended information is also suitable for recommending the bid advertisements. The target sample recommendation message obtained by the neural network of the embodiment of the application overcomes the problem that in the scheme for recommending the contracted advertisement in the prior art, the recommendation scheme cannot meet the maximum benefit of recommendation intermediaries because the dimension of the message recommendation model is the recommendation probability and the bid advertisement gives the recommendation value, so that the participation of the bid advertisement cannot be considered due to different dimensions.

It should be understood that, after determining the target sample recommendation message of each sample recommendation object, the recommendation results of all sample recommendation messages may be obtained, and the recommendation results may be which sample recommendation objects each sample recommendation message will be recommended to, and the number of times, cost, effect, etc. of the sample recommendation message recommendation, which are not specifically limited in the embodiments of the present application.

S104, determining a first difference between a target recommendation index and an actual recommendation index corresponding to each sample recommendation message, determining a second sequencing result of each sample recommendation message recommended to each sample recommendation object according to the first recommendation value, determining a second sequencing result of each sample recommendation object and a second difference between the first sequencing results, performing iterative training on the neural network model according to the first difference and the second difference until an iteration stop condition is met, and taking the neural network model after the iteration stop as a message recommendation model.

According to the embodiment of the application, iterative training is carried out on the neural network model to consider the difference of two dimensions, one is based on the first difference between the target recommendation index corresponding to each sample recommendation message and the actual recommendation index, the neural network model is trained by utilizing the first difference, the recommendation result determined on the basis of the neural network model can be enabled to meet the requirement of a recommendation subject as much as possible, the benefits of the recommendation subject and a recommendation intermediary are balanced, and the other is the second difference between two sequencing results, wherein the second sequencing result is obtained on the basis of the first recommendation value estimated by all recommendation subjects, the first sequencing result is obtained on the basis of the second recommendation value estimated by the model according to the message characteristics and the object characteristics, and because the output of the neural network model is the sequencing result, on one hand, the sequencing result can directly determine whether the recommendation message can be recommended (reflect the recommendation times), and on the other hand, the sequencing result is derived from the recommendation value (reflect the recommendation values), so that the scheme skillfully solves the defect that the existing recommendation model only considers one dimension, namely the recommendation probability or recommendation value is not suitable for a mixed scenario.

Based on the above embodiments, as an optional embodiment, the plurality of sample recommendation messages in the embodiment of the present application includes at least one first-class sample recommendation message and at least one second-class sample recommendation message, where the first-class sample recommendation message is a sample recommendation message that is ordered and recommended according to the recommendation number, and the second-class sample recommendation message is a sample recommendation message that is ordered and recommended according to the recommendation value.

Taking the advertisement putting scene as an example, the first category of sample recommended messages are contract advertisements, and the second category of sample recommended messages are bid advertisements.

The embodiment of the application further comprises target effect recommendation times reaching a first preset recommendation effect on the basis of comprising the target recommendation times aiming at the target recommendation indexes of the first class sample recommendation message. The recommending effect, namely an effect achieved after the sample recommending message is recommended to the sample recommending object, is exemplified by an advertisement putting scene, and common recommending effects comprise clicking, conversion and the like, wherein clicking refers to the action of clicking by the sample recommending object after the sample recommending message is displayed. Conversion refers to the act of reaching the web page indicated by the sample recommendation message and generating a purchase.

The recommendation effect can see if the recommended sample recommendation message attracts the sample recommendation object. The specific type of recommendation effect may be indicated by the recommendation body to the recommendation intermediary. For example, a target recommendation index of a certain recommendation subject for a certain first-class sample recommendation message is target recommendation frequency 10000, and target effect recommendation frequency 7000 of click effect is achieved, that is, the recommendation subject hopes that the first-class sample recommendation message can recommend 10000 times to different sample recommendation objects in a recommendation period, and hopes that 7000 times of 10000 times of recommendation can generate click behaviors for the sample recommendation objects.

It can be understood that the actual recommendation index of the first class sample recommendation message correspondingly includes the actual recommendation frequency and the actual effect recommendation frequency for achieving the first preset recommendation effect, and the subsequent embodiment of the present application will explain the method for obtaining the actual recommendation frequency and the actual effect recommendation frequency.

The target recommendation index of the second-class sample recommendation message according to the embodiment of the application can comprise target recommendation cost and target effect recommendation cost for achieving a second preset recommendation effect, wherein the target recommendation cost only considers the cost generated by the recommendation action, and the target effect recommendation cost further considers the cost generated when the recommendation achieves the second preset recommendation effect on the basis of the recommendation action. According to the embodiment of the application, different recommendation indexes are indicated for different types of sample recommendation messages, wherein for a first type of sample recommendation messages, the characteristics of the different types of sample recommendation messages can be learned by a neural network model under the consideration of the recommendation cost and the effect recommendation cost for generating the recommendation effect when the recommendation frequency and the effect recommendation frequency for generating the recommendation effect are considered for a second type of sample recommendation messages, and the accurate recommendation of the two types of sample recommendation messages is finally realized.

It may be appreciated that, in the embodiment of the present application, the first difference includes a difference between the target recommended number and the actual recommended number, a difference between the target effect recommended number and the actual effect recommended number, a difference between the target recommended cost and the actual recommended cost, and a difference between the target effect recommended cost and the actual effect recommended cost, which are 4 types of differences in total.

The embodiment of the application does not specifically limit whether the first preset recommended effect and the second preset recommended effect are the same.

It will be appreciated that if the actual recommendation index is better than the target recommendation index, it indicates that the recommendation entity obtains the expected benefit, and the recommendation intermediary has a partially ineffective benefit, if the target recommendation index is better than the actual recommendation index, it indicates that the recommendation entity does not obtain the expected benefit, and the recommendation intermediary needs to pay out a part of the benefit of the recommendation entity, so the neural network model of the present application is expected to reduce the difference between the actual recommendation index and the target recommendation index as much as possible during training.

On the basis of the foregoing embodiments, as an alternative embodiment, obtaining an actual recommendation index of each sample recommendation message according to a recommendation result of each sample recommendation message as a target sample recommendation message, including:

And regarding each first-class sample recommended message, taking the first-class sample recommended message as the number of target sample recommended messages and taking the number of target sample recommended messages as the actual recommended times. It can be understood that each sample recommendation message of the first category is regarded as a target sample recommendation message of one sample recommendation object, and therefore, the actual recommendation times can be obtained by counting the number of the sample recommendation messages of the first category as the target sample recommendation messages.

And obtaining an effect coefficient of the first-class sample recommendation message serving as a target sample recommendation message for the corresponding sample recommendation object for each first-class sample recommendation message, wherein the effect coefficient is used for representing the probability that the sample recommendation message is recommended to the sample object to reach a preset recommendation effect. It should be appreciated that the effect coefficients are predetermined and are not updated as the neural network model is trained.

And according to the effect coefficient of the corresponding sample recommended object when the first class of sample recommended information is used as the target sample recommended information, obtaining actual effect recommended times. Specifically, since the effect coefficients when the first-type sample recommendation message is used as the target sample recommendation message of the different sample recommendation objects are determined, the actual effect recommendation times can be obtained by accumulating the effect coefficients. For example, if a certain first-class sample recommendation message is used as the target sample recommendation message of the sample recommendation object 1 to the sample recommendation object 10000, the effect coefficient of the first-class sample recommendation message for the sample recommendation object 1 to the sample recommendation object 5000 is 0.3, and the effect coefficient of the first-class sample recommendation message for the sample recommendation object 5001 to the sample recommendation object 10000 is 0.4, the actual effect recommendation number is 35000.

For each second-class sample recommendation message, the embodiment of the application firstly determines the second-class sample recommendation message as a sample recommendation object of the target sample recommendation message and uses the second-class sample recommendation message as the target sample recommendation object. For example, if the template sample recommendation message of the sample recommendation object a is the second type sample recommendation message b, the sample recommendation object a is the target sample recommendation object of the second type sample recommendation message b.

After determining that the target sample recommends objects, the embodiment of the application further uses the second recommendation value immediately after the second recommendation value of the second class of sample recommendation messages as the reference value of the target sample recommends objects from the first sequencing result of the target sample recommends objects. In general, the number of target sample recommendation messages of each sample recommendation object is only 1, that is, the first sample recommendation message is ranked in the first ranking result, when one sample recommendation message is taken as the target sample recommendation message, one second sample recommendation message immediately after the second recommendation value of the message is the second sample recommendation message with highest value in the sample recommendation messages with failed recommendation, and when the second sample recommendation message with second class is not available, the sample recommendation message with second original rank is recommended, so that the second recommendation value immediately after the second recommendation value of the second class sample recommendation message can be taken as the reference value of the target sample recommendation object. And summing the reference values of all target sample recommended objects of the second-class sample recommended message to obtain the actual recommended cost of the second-class sample recommended message.

According to the embodiment of the application, for each second-class sample recommendation message, the effect coefficient of the second-class sample recommendation message on the corresponding sample recommendation object is obtained when the second-class sample recommendation message is taken as the target sample recommendation message, and according to the embodiment, the effect coefficient is used for representing the probability that the second-class sample recommendation message is recommended to the corresponding sample recommendation object to reach the second preset recommendation effect, so that the actual effect recommendation times of the second-class sample recommendation message to reach the second preset recommendation effect can be obtained by accumulating the effect coefficient of the second-class sample recommendation message on the corresponding sample recommendation object when the second-class sample recommendation message is taken as the target sample recommendation message.

The actual effect recommendation cost of the second type sample recommendation message reaching the second preset recommendation effect can be obtained by dividing the actual recommendation cost obtained in the embodiment by the actual effect recommendation times of the second type sample recommendation message, and it can be understood that the actual cost needs to be paid when the second type sample recommendation message described by the actual effect recommendation cost obtains the recommendation of reaching the second preset recommendation effect every time.

On the basis of the above embodiments, as an alternative embodiment, the neural network model includes a first sub-neural network model and a second sub-neural network model;

the second sub-neural network model is used for obtaining a first sequencing result of each sample recommendation message recommended to each sample recommendation object according to the second recommendation value of the sample recommendation object for each sample recommendation message according to each sample recommendation object.

That is, the second recommended value and the first ranking result are both obtained through the sub-neural network model, particularly when the first ranking result is obtained, the (second) sub-neural network model is used to complete ranking, rather than simply ranking based on the value, because from the relationship between the second recommended value and the first ranking result, if the first ranking result is obtained through the sub-neural network model, the second sub-neural network model is located behind the first sub-neural network model, and when the neural network model is trained through the back propagation algorithm, the training result of the second sub-neural network model can be transmitted to the first sub-neural network model, so that the training accuracy of the first sub-neural network model is higher.

Performing iterative training on the neural network model according to the first difference and the second difference until an iteration stop condition is reached, and taking the neural network model after the iteration stop as a message recommendation model, wherein the method comprises the following operations of repeatedly executing until the iteration stop condition is reached:

Referring to fig. 4, which is an exemplary illustration showing a training flow diagram of a message recommendation model according to an embodiment of the present application, as shown in the drawing, a target recommendation index, a message characteristic, and a bid (i.e., a first recommendation value) of each sample recommendation object for each sample recommendation message are obtained, m and n are positive integers greater than 1, based on a magnitude order of n first recommendation values of each sample recommendation object, a first recommendation result of each sample recommendation object is generated, a message characteristic of n sample recommendation messages and an object characteristic of m recommendation objects are input to a first sub-neural network model, thereby obtaining a second recommendation value of each sample recommendation message for each sample recommendation message, m in total, n second recommendation values of each sample recommendation object are input to a second sub-neural network model, a second ranking result of each sample recommendation object is obtained, the second ranking results record an order in which each sample recommendation message is recommended, a first recommendation result of each sample recommendation is based on a second ranking result of each sample recommendation object, and a first sub-neural network model is obtained, and a first sub-neural network is able to obtain a first recommendation index of each sample recommendation, and a first sub-neural network is able to obtain a first recommendation index of n actual samples, and a second recommendation index of each sample recommendation object is obtained, training the model parameters of the second sub-neural network according to the m groups of second differences.

On the basis of the above embodiments, as an alternative embodiment, the first sub-neural network model includes a recommended parameter layer and a value layer.

The recommendation parameter layer is used for recommending parameter layers to obtain initial recommendation values of the sample recommendation messages according to the message characteristics of the sample recommendation messages and effect values of preset popularization effects generated when the sample recommendation messages are recommended according to the message characteristics of the sample recommendation messages and object characteristics of a plurality of sample recommendation objects.

In an alternative embodiment, the recommended parameters layer may be a multi-layer perceptron (Multi Layer Perceptron, MLP), which has the advantage of producing a more accurate expression capability by constantly stacking hidden layers therein.

It should be noted that, the recommendation parameter layer is configured to obtain, on the one hand, an initial recommendation value of the sample recommendation message according to the message feature of the sample recommendation message, where it is understood that the initial recommendation value is a recommendation value of the sample recommendation message itself and is irrelevant to the sample recommendation object, and on the other hand, determine, on the other hand, an effect value of the sample recommendation message that generates a preset recommendation effect when recommending according to the message feature of the sample recommendation message and object features of the plurality of sample recommendation objects, where the effect value is learned by combining the message feature and the object features by the neural network model, and is continuously optimized through training.

The first sub-neural network model of the embodiment of the application obtains the second recommendation value not directly according to the message characteristics and the object characteristics, but firstly obtains the initial recommendation value of the sample recommendation message and the effect value for generating the preset popularization effect through the recommendation parameter layer, and can obtain the reference effect value of the sample recommendation object for generating the recommendation effect on the sample recommendation message through the effect value and the recommendation coefficient, and can obtain the second recommendation value of the sample recommendation object for each sample recommendation message based on the reference effect value and the initial recommendation value. That is, the second recommendation value obtained by the present application includes both the initial recommendation value of the sample recommendation message itself and the effect value of the sample recommendation object for generating the recommendation effect on the sample recommendation message, thereby defining the recommendation value of a recommendation message more accurately.

Referring to fig. 5, which illustrates a training method of a message recommendation model according to another embodiment of the present application, as shown in the drawing, the present application firstly obtains the message characteristics of n sample recommendation messages and the recommendation characteristics of m sample recommendation objects, matches the directional relation between the sample recommendation messages and the sample recommendation objects to obtain two graphs, performs feature aggregation according to the two graphs to perform feature expansion of the sample recommendation messages, so that the new message characteristics of each sample recommendation message are fused with the recommendation characteristics of each sample recommendation object, only needs to process the new message characteristics of the sample recommendation messages subsequently, improves the model training efficiency, inputs the new message characteristics of each sample recommendation message to a recommendation parameter layer of an MLP structure of a first sub-neural network model (not shown in the drawing), obtains the algorithm parameters of each sample recommendation message, the algorithm parameters of the embodiment of the present application include initial recommendation values and effect values, the value layer (not shown in the drawing) is based on the algorithm parameters and recommendation coefficients, can obtain a second value of each sample recommendation message for each sample recommendation message, differential network (Differential Sorting Network) is used for obtaining a recommendation value of each sample recommendation message, and a recommendation value of each sample recommendation is obtained according to a second sample recommendation result of the actual ranking function, and a recommendation result is obtained by combining the recommendation value of each sample recommendation object with a second sample recommendation result of each sample recommendation object, and a recommendation value of the actual ranking function is obtained according to the actual result, and updating parameters of the recommended parameter layer through a first loss function value, obtaining a second loss value of a second loss function according to a second difference between a first sorting result and a second sorting result of the sample recommended object, updating parameters of a second sub-neural network model through the second loss value, updating parameters of the recommended parameter layer through a back propagation algorithm until the first loss function and the second loss function are converged, ending iterative training, and obtaining the message recommended model.

The embodiment of the application also provides a message recommending method, as shown in fig. 6, comprising the following steps:

s201, obtaining message characteristics of a plurality of recommended messages and object characteristics of at least one recommended object;

s202, inputting message characteristics of a plurality of recommended messages and object characteristics of a plurality of recommended objects into a message recommendation model, and obtaining a third sequencing result of each recommended message which is determined by the message recommendation model and recommended to each recommended object;

s203, determining a target recommendation message recommended to each recommended object based on the third sequencing result of the recommended object, and recommending the target recommendation message to the recommended object.

Referring to fig. 7, a system architecture diagram of a message recommendation system according to an embodiment of the present application is exemplarily shown, and as shown in the figure, an advertisement delivery system includes a streaming media server, an object feature server, an advertisement delivery server, and an advertisement database.

The streaming media server in the embodiment of the application can provide streaming media service for the terminal, in the social field, the streaming media service provided by the streaming media server can be social service, such as providing social dynamics, social objects and the like, in the game field, the streaming media service provided by the streaming media server can be game service, such as providing downloadable games, providing game running environment and the like, and in the video playing field, the streaming media server provided by the streaming media server can be video playing, video uploading service and the like.

The terminal is connected with the streaming media server to apply for streaming media service, the terminal can carry some characteristic information, such as terminal model, account information and the like, when the terminal applies for streaming media service, the information is received by the streaming media server and then is sent to the advertisement putting server as terminal information, the advertisement putting server obtains the latest object characteristic of the terminal information (corresponding to an access object) from the object characteristic server according to the terminal information, because the advertisement putting server can put a lot of advertisements, the advertisement range can be reduced firstly, the advertisement putting server of the embodiment can determine the advertisement screening range according to the terminal information and the type of the streaming media service provided by the streaming media server, then applies the advertisement characteristic server for the message characteristic of a plurality of advertisements (namely recommended messages) in the range, the advertisement characteristic server inputs the latest object characteristic and the message characteristic of the plurality of recommended messages into the message recommendation model trained by the embodiment of the application, determines the target recommended messages (namely target advertisements) to be recommended to the terminal, the advertisement characteristic server obtains the target message from the advertisement database, the advertisement characteristic server can integrate the target message information with the streaming media server, and the target recommended information is updated to the advertisement characteristic database when the advertisement characteristic is updated by the advertisement server, and the advertisement characteristic is updated by the advertisement server.

The embodiment of the application provides a training device for a message recommendation model, as shown in fig. 8, the training device for the message recommendation model may include: a sample data acquisition module 101, a ranking module 102, an index determination module 103, and an iterative training module 104, wherein,

a sample data obtaining module 101, configured to obtain message characteristics of a plurality of sample recommendation messages, object characteristics of a plurality of sample recommendation objects, a target recommendation index of the sample recommendation message indicated by a recommendation subject of each sample recommendation message to a recommendation intermediary, and a first recommendation value of each sample recommendation object for the sample recommendation message;

a ranking module 102, configured to input the message characteristics of the plurality of sample recommendation messages and the object characteristics of the plurality of sample recommendation objects to the neural network model, obtain a second recommendation value of the sample recommendation object for each sample recommendation message determined by the neural network model for each sample recommendation object, and a first ranking result of each sample recommendation message recommended to the sample recommendation object determined based on the second recommendation value;

an index determining module 103, configured to determine, for each sample recommended object, a target sample recommended message recommended to the sample recommended object based on a first sorting result of the sample recommended object, and obtain an actual recommendation index of each sample recommended message according to a recommendation result of each sample recommended message as a target sample recommended message;

The iterative training module 104 is configured to determine a second sorting result of each sample recommendation message recommended to each sample recommendation object according to the first recommendation value, determine a first difference between a target recommendation index and an actual recommendation index corresponding to each sample recommendation message and a second difference between a second sorting result and the first sorting result of each sample recommendation object, perform iterative training on the neural network model according to the first difference and the second difference until an iteration stop condition is reached, and use the neural network model after the iteration stop as a message recommendation model.

The device of the embodiment of the present application may execute the training method of the message recommendation model provided by the embodiment of the present application, and its implementation principle is similar, and actions executed by each module in the device of each embodiment of the present application correspond to steps in the training method of the message recommendation model of each embodiment of the present application, and detailed functional descriptions of each module of the device may be referred to the descriptions in the corresponding methods shown in the foregoing, which are not repeated herein.

The embodiment of the application provides a message recommending device, as shown in fig. 9, which comprises a pending data obtaining module 201, a sorting determining module 202 and a recommending module 203, specifically:

A data to be processed obtaining module 201, configured to obtain message characteristics of a plurality of recommended messages and object characteristics of a plurality of recommended objects;

a ranking determining module 202, configured to input the message characteristics of the plurality of recommended messages and the object characteristics of the plurality of recommended objects into a message recommendation model, and obtain a first ranking result of each recommended message recommended to the recommended object for each recommended object determined by the message recommendation model;

the recommending module 203 is configured to determine, for each recommended object, a target recommending message recommended to the recommended object based on the first sorting result of the recommended object, and recommend the target recommending message to the recommended object.

The device of the embodiment of the present application may execute the message recommendation method provided by the embodiment of the present application, and its implementation principle is similar, and actions executed by each module in the device of each embodiment of the present application correspond to steps in the message recommendation method of each embodiment of the present application, and detailed functional descriptions of each module of the device may be referred to in the corresponding method shown in the foregoing, which is not repeated herein.

The embodiment of the application provides an electronic device, which comprises a memory, a processor and a computer program stored on the memory, wherein the processor executes the computer program to realize a training method of a message recommendation model or the steps of the message recommendation method, and compared with the related technology, the method can realize the following steps: the method comprises the steps of carrying out iterative training on a neural network model to consider differences of two dimensions, wherein one is based on a first difference between a target recommendation index and an actual recommendation index corresponding to each sample recommendation message, training the neural network model by utilizing the first difference, enabling recommendation results determined on the basis of the neural network model to meet the requirements of recommendation subjects as much as possible, balancing benefits of the recommendation subjects and recommendation intermediaries, and the other is a second difference between two sorting results, wherein the second sorting results are obtained on the basis of first recommendation values evaluated by all recommendation subjects, the first sorting results are obtained on the basis of second recommendation values evaluated by the model according to message characteristics and object characteristics, and because the output of the neural network model is the sorting results, on one hand, the sorting results can directly determine whether recommendation messages can be recommended (reflect recommendation times), and on the other hand, the sorting results are derived from recommendation values (reflect recommendation values), so that the scheme skillfully solves the defects that the existing recommendation model only takes one dimension, namely recommendation probability or recommendation value, is not suitable for mixed-ranking scenes.

In an alternative embodiment, there is provided an electronic device, as shown in fig. 10, the electronic device 4000 shown in fig. 10 includes: a processor 4001 and a memory 4003. Wherein the processor 4001 is coupled to the memory 4003, such as via a bus 4002. Optionally, the electronic device 4000 may further comprise a transceiver 4004, the transceiver 4004 may be used for data interaction between the electronic device and other electronic devices, such as transmission of data and/or reception of data, etc. It should be noted that, in practical applications, the transceiver 4004 is not limited to one, and the structure of the electronic device 4000 is not limited to the embodiment of the present application.

The processor 4001 may be a CPU (Central Processing Unit ), general purpose processor, DSP (Digital Signal Processor, data signal processor), ASIC (Application Specific Integrated Circuit ), FPGA (Field Programmable Gate Array, field programmable gate array) or other programmable logic device, transistor logic device, hardware components, or any combination thereof. Which may implement or perform the various exemplary logic blocks, modules and circuits described in connection with this disclosure. The processor 4001 may also be a combination that implements computing functionality, e.g., comprising one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

Bus 4002 may include a path to transfer information between the aforementioned components. Bus 4002 may be a PCI (Peripheral Component Interconnect, peripheral component interconnect standard) bus or an EISA (Extended Industry Standard Architecture ) bus, or the like. The bus 4002 can be divided into an address bus, a data bus, a control bus, and the like. For ease of illustration, only one thick line is shown in fig. 10, but not only one bus or one type of bus.

Memory 4003 may be, but is not limited to, ROM (Read Only Memory) or other type of static storage device that can store static information and instructions, RAM (Random Access Memory ) or other type of dynamic storage device that can store information and instructions, EEPROM (Electrically Erasable Programmable Read Only Memory ), CD-ROM (Compact Disc Read Only Memory, compact disc Read Only Memory) or other optical disk storage, optical disk storage (including compact discs, laser discs, optical discs, digital versatile discs, blu-ray discs, etc.), magnetic disk storage media, other magnetic storage devices, or any other medium that can be used to carry or store a computer program and that can be Read by a computer.

The memory 4003 is used for storing a computer program for executing an embodiment of the present application, and is controlled to be executed by the processor 4001. The processor 4001 is configured to execute a computer program stored in the memory 4003 to realize the steps shown in the foregoing method embodiment.

Embodiments of the present application provide a computer readable storage medium having a computer program stored thereon, which when executed by a processor, implements the steps of the foregoing method embodiments and corresponding content.

The embodiment of the application also provides a computer program product, which comprises a computer program, wherein the computer program can realize the steps and corresponding contents of the embodiment of the method when being executed by a processor.

The terms "first," "second," "third," "fourth," "1," "2," and the like in the description and in the claims and in the above figures, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged where appropriate, such that the embodiments of the application described herein may be implemented in other sequences than those illustrated or otherwise described.

It should be understood that, although various operation steps are indicated by arrows in the flowcharts of the embodiments of the present application, the order in which these steps are implemented is not limited to the order indicated by the arrows. In some implementations of embodiments of the application, the implementation steps in the flowcharts may be performed in other orders as desired, unless explicitly stated herein. Furthermore, some or all of the steps in the flowcharts may include multiple sub-steps or multiple stages based on the actual implementation scenario. Some or all of these sub-steps or phases may be performed at the same time, or each of these sub-steps or phases may be performed at different times, respectively. In the case of different execution time, the execution sequence of the sub-steps or stages can be flexibly configured according to the requirement, which is not limited by the embodiment of the present application.

The foregoing is only an optional implementation manner of some implementation scenarios of the present application, and it should be noted that, for those skilled in the art, other similar implementation manners based on the technical ideas of the present application are adopted without departing from the technical ideas of the scheme of the present application, which also belongs to the protection scope of the embodiments of the present application.

Claims

1. A method for training a message recommendation model, comprising:

inputting the message characteristics of the plurality of sample recommendation messages and the object characteristics of the plurality of sample recommendation objects into a neural network model, and obtaining a second recommendation value of each sample recommendation message of each sample recommendation object determined by the neural network model and a first sequencing result of each sample recommendation message recommended to the sample recommendation object determined based on the second recommendation value, for each sample recommendation object;

determining a second sorting result of each sample recommendation message recommended to each sample recommendation object according to the first recommendation value, determining a first difference between a target recommendation index and an actual recommendation index corresponding to each sample recommendation message and a second difference between a second sorting result and the first sorting result of each sample recommendation object, performing iterative training on the neural network model according to the first difference and the second difference until an iteration stopping condition is reached, and taking the neural network model after iteration stopping as a message recommendation model.

2. The training method of claim 1, wherein the plurality of sample recommendation messages include at least one first-class sample recommendation message and at least one second-class sample recommendation message, the first-class sample recommendation message being a sample recommendation message ordered and recommended according to a recommendation number, the second-class sample recommendation message being a sample recommendation message ordered and recommended according to a recommendation value;

3. The training method according to claim 2, wherein the obtaining the actual recommendation index of each sample recommendation message based on the recommendation result of each sample recommendation message as the target sample recommendation message comprises:

and for each first-class sample recommendation message, obtaining an effect coefficient of the first-class sample recommendation message on the corresponding sample recommendation object when the first-class sample recommendation message is used as a target sample recommendation message, wherein the effect coefficient is used for representing the probability that the sample recommendation message is recommended to the sample object to reach a preset recommendation effect, and obtaining the actual effect recommendation times according to the effect coefficient of the first-class sample recommendation message on the corresponding sample recommendation object when the first-class sample recommendation message is used as the target sample recommendation message.

4. The training method according to claim 2, wherein the obtaining the actual recommendation index of each sample recommendation message based on the recommendation result of each sample recommendation message as the target sample recommendation message comprises:

5. The training method of claim 4, wherein the obtaining the actual recommendation index of each sample recommendation message based on the recommendation result of each sample recommendation message as the target sample recommendation message comprises:

6. The training method of any of claims 1-5, wherein the neural network model comprises a first sub-neural network model and a second sub-neural network model;

The second sub-neural network model is used for acquiring a first sequencing result of each sample recommendation message recommended to each sample recommendation object according to the second recommendation value of the sample recommendation object for each sample recommendation message for each sample recommendation object;

performing iterative training on the neural network model according to the first difference and the second difference until an iteration stopping condition is reached, and taking the neural network model after the iteration stopping as a message recommending model, wherein the method comprises the following operations of repeatedly executing until the iteration stopping condition is reached:

7. The training method of claim 6, wherein the first sub-neural network model includes a recommended parameters layer and a value layer;

The recommendation parameter layer is used for obtaining initial recommendation values of the sample recommendation messages according to the message characteristics of the sample recommendation messages and obtaining effect values of preset popularization effects when the sample recommendation messages are recommended according to the message characteristics of the sample recommendation messages and the object characteristics of the plurality of sample recommendation objects;

the value layer is used for obtaining a recommendation coefficient of each sample recommendation message for each sample recommendation object, and obtaining a second recommendation value of each sample recommendation message for each sample recommendation object according to an initial recommendation value of the sample recommendation message, an effect value of a preset popularization effect generated when the sample recommendation message is recommended and a recommendation coefficient of the sample recommendation message for each sample recommendation object.

8. A message recommendation method, comprising:

wherein the message recommendation model is trained based on the method of any one of claims 1-7.

9. A training device for a message recommendation model, comprising:

the sorting module is used for inputting the message characteristics of the plurality of sample recommendation messages and the object characteristics of the plurality of sample recommendation objects into a neural network model, and obtaining a second recommendation value of the sample recommendation object for each sample recommendation message determined by the neural network model and a first sorting result of each sample recommendation message recommended to the sample recommendation object, wherein the first sorting result is determined based on the second recommendation value;

and the iterative training module is used for determining a second sorting result of each sample recommendation message recommended to each sample recommendation object according to the first recommendation value, determining a first difference between a target recommendation index and an actual recommendation index corresponding to each sample recommendation message and a second difference between a second sorting result and the first sorting result of each sample recommendation object, performing iterative training on the neural network model according to the first difference and the second difference until an iterative stopping condition is reached, and taking the neural network model after iterative stopping as a message recommendation model.

10. A message recommendation device, comprising:

the ordering determining module is used for inputting the message characteristics of the plurality of recommended messages and the object characteristics of the plurality of recommended objects into the message recommending model, and obtaining a first ordering result of each recommended message which is determined by the message recommending model and recommended to each recommended object;

wherein the message recommendation model is trained based on the apparatus of claim 9.

11. An electronic device comprising a memory, a processor and a computer program stored on the memory, characterized in that the processor executes the computer program to carry out the steps of the method according to any one of claims 1-8.

12. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any of claims 1-8.

13. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any of claims 1-8.