CN113672797A

CN113672797A - Content recommendation method and device

Info

Publication number: CN113672797A
Application number: CN202010401115.7A
Authority: CN
Inventors: 张梦一; 王松; 郭腾蛟; 李邦鹏; 贺旭; 陈功
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2020-05-13
Filing date: 2020-05-13
Publication date: 2021-11-19

Abstract

The application relates to the technical field of computers, in particular to a content recommendation method and device, which are used for improving the content recommendation efficiency and effect. The method comprises the following steps: determining a user group to which a target user belongs; determining content evaluation index information corresponding to the user group, wherein the content evaluation index information is obtained by calculation according to historical behavior data of users in the user group aiming at recommended content; evaluating each content to be recommended of the target user according to the content evaluation index information; and determining recommended content recommended to the target user from all the contents to be recommended according to the evaluation result.

Description

Content recommendation method and device

Technical Field

The present application relates to the field of computer technologies, and in particular, to a content recommendation method and apparatus.

Background

With the development of the internet, online advertisements have become a mainstream advertisement delivery mode. Media owners insert advertisement slots to display online advertisements, and while an advertisement trading platform converts user traffic into cash revenue, user experience is often considered to ensure good development. Because users have different sensitivities and needs for advertisements, the emphasis on recommending advertisements to users is different. Personalized advertisement recommendation is needed for different users, and under the condition that the fluctuation of advertisement categories is not large, the balance between advertisement income and user experience is achieved.

In the existing recommended advertisement selection process, on-line test is generally performed for each user experience index, and corresponding advertisements are selected for different users to be recommended. The method needs more experimental data, and particularly when the user experience indexes are more, the required experimental data amount is huge, the calculation process is complex, and the pressure of flow is higher.

Disclosure of Invention

The embodiment of the application provides a content recommendation method and device, which are used for improving the content recommendation efficiency and effect.

According to a first aspect of embodiments of the present application, there is provided a content recommendation method, including:

determining a user group to which a target user belongs;

determining content evaluation index information corresponding to the user group, wherein the content evaluation index information is obtained by estimation according to historical behavior data of users in the user group aiming at recommended content;

evaluating each content to be recommended of the target user according to the content evaluation index information;

and determining recommended contents recommended to the target user from the contents to be recommended according to the evaluation result.

According to a second aspect of embodiments of the present application, there is provided a content recommendation apparatus, the apparatus including:

the grouping unit is used for determining the user group to which the target user belongs;

the index unit is used for determining content evaluation index information corresponding to the user group, and the content evaluation index information is calculated according to historical behavior data of users in the user group aiming at recommended content;

the evaluation unit is used for evaluating each content to be recommended of the target user according to the content evaluation index information;

and the determining unit is used for determining recommended contents recommended to the target user from all the contents to be recommended according to the evaluation result.

In an optional embodiment, the content evaluation index information includes at least two content evaluation indexes and a weight of each content evaluation index;

and calculating the weight corresponding to each content evaluation index according to sample content obtained from the historical behavior data of the recommended content by the users in the user group.

In an alternative embodiment, the evaluation unit is specifically configured to:

aiming at each content to be recommended, calculating a recommendation score of the content to be recommended by utilizing an estimation value of a content evaluation index of the content to be recommended and a weight corresponding to each content evaluation index;

and the estimation value of the content evaluation index of the content to be recommended is obtained by estimation according to the content characteristic value of the content to be recommended and the user characteristic value of the target user.

In an optional embodiment, the evaluation unit is specifically configured to determine an evaluation value of a content evaluation index of the content to be recommended according to the following manner:

inputting the content characteristic value of the content to be recommended and the user characteristic value of the target user into a trained deep neural network model to obtain an estimated value of a content evaluation index of the content to be recommended;

and the deep neural network model is trained according to the content characteristic value and the user characteristic value of the sample content and the interaction behavior data of the user aiming at the sample content to obtain corresponding model parameters.

In an optional embodiment, the evaluation unit is specifically configured to determine a weight corresponding to the content evaluation index according to the following manner:

at least obtaining an estimation value of a content evaluation index of recommended content and a final objective function corresponding to the user group, wherein the final objective function comprises a weight parameter of the content evaluation index of the recommended content;

determining a gradient of a weight parameter of the final objective function for the content evaluation index;

and performing iterative computation on the gradient according to a gradient descent method by using the estimated value of the content evaluation index of the recommended content, and determining the weight of the corresponding content evaluation index when the difference between two adjacent iterations is smaller than a preset threshold or reaches the iteration times.

In an alternative embodiment, the evaluation unit is specifically configured to determine the evaluation value of the content evaluation index of the recommended content according to the following manner:

inputting the content characteristic value of the recommended content and the user characteristic value of the recommended content into a trained deep neural network model, and calculating to obtain an estimated value of a content evaluation index of the recommended content;

In an optional embodiment, the evaluation unit is specifically configured to determine a final objective function corresponding to the user group according to the following manner:

determining a constraint condition and an initial objective function corresponding to the user group;

and determining the final objective function according to the constraint condition and the initial objective function.

combining the constraints and the initial objective function into a transition objective function;

and converting the non-differentiable items in the transition objective function into differentiable items to obtain the final objective function.

In an optional embodiment, the apparatus further comprises a filter unit for:

determining all relevant contents corresponding to the user groups according to the content evaluation index information corresponding to the user groups;

and determining the content to be recommended from all the related content according to the filtering rule.

According to a third aspect of embodiments of the present application, there is provided a computing device comprising at least one processor, and at least one memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the content recommendation method provided by embodiments of the present application.

According to a fourth aspect of the embodiments of the present application, there is provided a storage medium storing computer instructions, which, when run on a computer, cause the computer to perform the steps of the content recommendation method provided by the embodiments of the present application.

According to the method and the device, a plurality of user groups are set based on a set rule, and for each user group, the content evaluation index information of the user group is estimated and obtained according to the historical behavior data of the users in the user group for the recommended content. In the process of recommending content to a target user on line, aiming at the fact that the determined target user can match a plurality of contents to be recommended, recommended contents need to be selected from the contents to be recommended and pushed to the user. The specific recommendation method is to determine a user group to which a target user belongs and obtain content evaluation index information corresponding to the user group. And evaluating each content to be recommended of the target user according to the content evaluation index information, and determining recommended content recommended to the target user from all the contents to be recommended according to the evaluation result. In the embodiment of the application, the content evaluation index information corresponds to the user group and is obtained by calculation according to the historical behavior data of the users in the user group, so that compared with the method of respectively calculating by using single user, the calculation process is simplified, and the calculation data amount is reduced. And when the recommended content of the target user is determined, only the user group of the target user needs to be determined, namely the user group content evaluation index information of the user group can be directly used for evaluating the plurality of contents to be recommended respectively, so that the recommended content is selected from the plurality of contents to be recommended. Therefore, the technical scheme of the application has good performance on improving the efficiency and the effect of content recommendation.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present application.

Fig. 1 shows a system architecture diagram of a content recommendation system in an embodiment of the present application;

fig. 2 shows a flowchart of a content recommendation method in an embodiment of the present application;

FIG. 3 is a schematic diagram illustrating a structure of an optimization algorithm model in an embodiment of the present application;

FIG. 4 is a flow chart illustrating a method for advertisement recommendation in an exemplary embodiment of the present application;

fig. 5 is a block diagram showing a configuration of a content recommendation apparatus according to an embodiment of the present application;

fig. 6 is a block diagram illustrating a server according to an embodiment of the present application.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are some embodiments, but not all embodiments, of the technical solutions of the present application. All other embodiments obtained by a person skilled in the art without any inventive step based on the embodiments described in the present application are within the scope of the protection of the present application.

The terms "first" and "second" in the description and claims of the present invention and the above-described drawings are used for distinguishing between different objects and not for describing a particular order. Furthermore, the term "comprises" and any variations thereof, which are intended to cover non-exclusive protection. For example, a process, method, system, article, or apparatus that comprises a list of steps or elements is not limited to only those steps or elements listed, but may alternatively include other steps or elements not listed, or inherent to such process, method, article, or apparatus.

Some concepts related to the embodiments of the present application are described below.

1. Artificial intelligence

Artificial Intelligence (AI) is a theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human Intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a manner similar to human intelligence. Artificial intelligence is the research of the design principle and the realization method of various intelligent machines, so that the machines have the functions of perception, reasoning and decision making. The artificial intelligence technology mainly comprises a computer vision technology, a voice processing technology, machine learning/deep learning and other directions.

2. Machine learning

Machine learning is a multi-field cross discipline, and relates to a plurality of disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and the like. The special research on how a computer simulates or realizes the learning behavior of human beings so as to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve the performance of the computer. Machine learning is the core of artificial intelligence, is the fundamental approach for computers to have intelligence, and is applied to all fields of artificial intelligence. Machine learning generally includes techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and the like.

3. Cloud technology

Cloud technology refers to a hosting technology for unifying serial resources such as hardware, software, network and the like 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 a general term of network technology, information technology, integration technology, management platform technology, application technology and the like applied in a Cloud computing business model, can form a resource pool, is used as required, and is flexible and convenient. Cloud computing technology will become an important support. Background services of the technical network system require a large amount of computing and storage resources, such as video websites, picture-like websites and more web portals. With the high development and application of the internet industry, each article may have its own identification mark and needs to be transmitted to a background system for logic processing, data in different levels are processed separately, and various industrial data need strong system background support and can only be realized through cloud computing.

4. Online advertisement, media owner, advertiser and advertisement trading platform

Online advertising, also known as internet advertising, refers to advertising placed on advertising spots (e.g., WeChat circle of friends, public post, news software, etc.) on the internet platform.

Media owners refer to entities that own the internet platform (e.g., WeChat friend circles, public numbers, news software, etc.) and generally have a large user access volume (also referred to as user traffic) and wish to convert the user traffic into cash revenue so they insert advertising spots into the platform.

The advertiser refers to an entity which displays the advertisement of the advertiser through an advertisement space of an internet platform.

An ad exchange (ADX) refers to an entity that connects a media host and an advertiser and that places the advertiser's ad on an ad spot provided by the media host.

5、CPM、eCPM、pCTR

CPM (Cost Per mile) refers to the Cost that an advertiser needs to pay after an advertisement is shown to a thousand visiting users on an internet platform.

eCPM (effective Cost Per Mile, thousand show revenue) is the advertising revenue that can be obtained Per thousand shows, reflecting the profitability of the platform, and can be regarded as an effective estimate for CPM.

pCTR (predicted Click Through Rate) is the probability that an advertisement is clicked after being delivered under a certain situation, which is predicted by the online advertising system.

6. Multi-objective optimization

Multi-objective optimization, also known as multi-objective planning or pareto optimization, is a field of multi-objective decision-making, and when there is a trade-off between two or more conflicting objectives, a balance point needs to be found to take an optimal decision. When one of the primary objectives is used as an objective function, the remaining objectives can be used as constraints to represent variables of the decision-making scheme, thereby imposing a limit on the decision-making scheme.

7. Relaxation (relax)

In the field of optimization, for the problem that the existence of an immutable or discontinuous function causes high solution complexity, other differentiable or continuous functions can be adopted to replace the original function, and the process is called relaxation. The optimization problem after relaxation is easier to solve, and the optimal solution can be seen as an approximation of the optimal solution of the original problem.

8. Content evaluation index, content evaluation index information, and weight of content evaluation index

The advertisement with the optimal comprehensive advertising income and user experience is selected from the advertisements to be recommended and recommended to the user, each advertisement to be recommended can be scored according to set indexes, and the advertisement with the highest score is determined from all the advertisements to be recommended and recommended to the user. The content evaluation index in the embodiment of the application corresponds to the setting index and is used for representing the characteristics of the advertisement and the experience of the user for the advertisement. The content evaluation index information includes content evaluation indexes of the advertisement and weights of each content evaluation index, and the content evaluation index information corresponds to the user groups, that is, for users in the same user group, the content evaluation indexes and the corresponding weights are the same. The content rating indicators may comprise two types of indicators, relating on the one hand to the characteristics of the advertisement itself, such as eCPM, etc., and on the other hand to the user's interaction behavior with respect to the advertisement, such as the approval rate, the negative feedback rate, etc. For the former content rating measure, the value of which is only related to the characteristics of the advertisement itself, e.g., eCPM, can be estimated directly from the characteristics of the advertisement, typically by evaluation by the advertiser. The latter content evaluation index has a value related not only to the characteristics of the advertisement itself but also to the user interacting with the advertisement, for example, a negative feedback rate, and therefore can be predicted by using a machine learning model. That is, the values of the content evaluation index are all estimated values in the present application.

The weight of the content rating index reflects the influence of each content rating index on whether the advertisement is recommended, and the larger the weight is, the more important the content rating index is. In the embodiment of the application, the weight of each content evaluation index is calculated by using a multi-objective optimization method, and the calculated weight corresponds to a user group, namely the weight values of the content evaluation indexes of different advertisements corresponding to the same user group are the same. And the estimation values of the content evaluation indexes are different among different advertisements, so that each advertisement can be scored by utilizing the weight value of the content evaluation index and the estimation value of the content evaluation index, and whether the advertisement is worth recommending or not can be determined according to the obtained score.

The basic concept of the present application is described below.

In the related art, generally, the balance between different interactive behaviors of the user and recommended content is mainly considered for recommending the content such as advertisements to the user. Taking the information flow scene as an example, the optimization target is the click rate and the like rate of the recommended content. Therefore, it is necessary to score recommended contents based on the click rate and the like. In the model for predicting the recommended content score, the click rate and the like rate are positive samples, but different weights are set, so that the final score is influenced. However, the weights are generally chosen based on empirical values or adjusted by means of an on-line AB test.

The concept of the AB test is derived from a biomedical double-blind test in which patients are randomly divided into two groups, and blindly given placebo and test medications, respectively, and after a period of experimentation, the performance of the two groups of patients is compared to see if there is a significant difference, thereby determining if the test medications are effective. The AB test also employs a similar concept: and respectively recommending the advertisements to a plurality of users for experiment in the same time dimension, collecting the interaction behavior data of each user aiming at each advertisement, and finally analyzing and evaluating the best advertisement.

The AB test is a process of iterative optimization, and its basic steps can be divided into:

step 1, setting a project target, namely an AB test target, for measuring the quality of each advertisement;

step 2, designing an optimized iterative development scheme to determine various weight combinations;

step 3, determining a plurality of advertisements to be recommended and the distribution ratio of each advertisement to be recommended;

step 4, opening the on-line flow according to the shunting proportion for testing;

step 5, collecting experimental data to judge effectiveness and effect;

and 6, determining a weight combination according to a test result, adjusting the shunt ratio to continue the test or continuing the optimization iteration scheme of the step 2 to redevelop the online test under the condition that the test effect is not achieved.

In the above-described online AB test, each set of weights needs to correspond to one set of online experiments. When content evaluation indexes (such as click rate, approval rate, comment rate, negative feedback rate, sharing rate and the like) are large, the weight dimensionality is increased, the required experiment number is exponentially increased, and the waste of flow is caused.

Based on this, the embodiment of the application directly determines the corresponding content evaluation index information for each group according to the user group to which the target user belongs, so that each content to be recommended of the target user in the group can be evaluated according to the content evaluation index information, and further, the recommended content of the target user is determined from all the content to be recommended according to the evaluation result. Therefore, the trial and error process of a plurality of groups of on-line experiments is avoided, the data volume is reduced, and the calculation difficulty is reduced.

Further, the recommended content is selected from the contents to be recommended according to the characteristics of the user groups, not only according to the personal characteristics of the user, but also according to the grouping characteristics related to the user, so that the recommended content expands the range of the recommended content, enhances the user experience and improves the recommendation effect on the basis of ensuring the relevance of the recommended content to the user individuals.

In the embodiment of the application, the recommended content is determined in a mode of scoring the content to be recommended. The specific scoring mode is that for each content to be recommended, the weights and the estimated values of all content evaluation indexes of the content to be recommended are weighted, so that a score is obtained. And then sequencing all the contents to be recommended according to the scores, and recommending the contents to be recommended with the highest scores as recommended contents to the user.

On one hand, the estimation value of the content evaluation index of the content to be recommended can be obtained through calculation, and calculation is performed according to the content characteristic value of the content to be recommended and the user characteristic value of the target user, for example, for the target user, the user characteristic value of the target user and the content characteristic value of one content to be recommended are input into a trained deep neural network model, so that the estimation value of the content evaluation index of the content to be recommended is obtained, and the deep neural network model is trained according to the content characteristic value of the sample content, the user characteristic value and interaction behavior data of the user for the sample content. That is to say, in the embodiment of the present application, the estimation value of the content evaluation index of each content to be recommended is estimated according to the sample content.

On the other hand, the weight of the content evaluation index of the content to be recommended can be obtained through calculation of an optimization algorithm model, each user group corresponds to the weight of one group of content evaluation indexes, and the weight can be obtained through calculation in advance and stored in a database. For a user group, a final objective function corresponding to the user group is obtained, and an estimated value of a content evaluation index of recommended content is obtained. And the final objective function is a function containing weight parameters, and the estimation value of the content evaluation index of the recommended content is also obtained by the estimation of the trained deep neural network model. And performing iterative computation according to a multi-objective optimization algorithm by using the final objective function and the estimated value of the content evaluation index of the recommended content to determine an optimal weight combination as the weight of the user group.

Generally, the selection of recommended content depends on content evaluation indexes needing to be optimized in a real scene, and because different content evaluation indexes are often in a restrictive relationship, for example, for a user with low interaction frequency with an advertisement, the interaction behavior frequency of the user, such as click rate, praise rate and the like, needs to be improved, so that the profit of an advertisement trading platform can be properly sacrificed, the pertinence of advertisement recommendation is improved, and a better advertisement display effect is achieved. Therefore, in the embodiment of the invention, when the weight of the same user group is calculated, the optimized targets include a plurality of targets, the optimized sequence among the targets is determined, one target is used as a main target to construct a target function, and the other targets are used as secondary targets to construct constraint conditions, so that the aim of balancing various relationships is fulfilled.

In addition, for different user groups, because the optimization targets are different, preferably, the target functions of different user groups are set to be different, and the constraint conditions are not completely the same, so that the weights of the content evaluation indexes are different for different user groups, and further, the content recommendation is more targeted.

In this way, for the target user, the user group of the target user is first specified, and the weight of the content evaluation index corresponding to the user group is acquired. And determining a plurality of contents to be recommended matched with the target user, and calculating an estimation value of a content evaluation index of the contents to be recommended aiming at each content to be recommended. And then, calculating the score of each content to be recommended according to the estimation value of the content evaluation index of the content to be recommended and the weight of the content evaluation index. And finally, sequencing the plurality of contents to be recommended according to the scores, and selecting the contents to be recommended with the highest scores as recommended contents to be sent to the user.

In addition, after the recommended content is sent to the user, interaction and feedback data of the user on the recommended content can be collected and input into the deep neural network model and the optimization algorithm model as sample content for training and updating of the two models.

In the embodiment of the application, the multi-objective optimization algorithm may adopt optimization algorithms such as a simulated annealing algorithm, a genetic algorithm, an ant colony algorithm, and the like to calculate the weight of the content evaluation index. Preferably, the embodiment of the application combines the relaxation variables with the gradient descent algorithm to perform iterative optimization, and outputs the optimal solution of the weight.

The relaxation variables and gradient descent algorithm are described in detail below.

If the constraints of the linear programming model are all smaller than type, then M non-negative relaxation variables can be introduced by the normalization process. The introduction of relaxation variables is often to facilitate solution over a larger feasible domain. If the value is 0, the state converges to the original state, and if the value is larger than zero, the constraint is relaxed.

In particular, the study of the linear programming problem is based on a standard type. Thus, for a given mathematical model of a non-standard type linear programming problem, it needs to be normalized. Generally, for different forms of linear programming models, some methods can be used to normalize them. When the constraint condition is less than the type of linear programming problem, a new non-negative variable can be added (or subtracted) to the left of the inequality, and the new non-negative variable can be converted into an equation. This newly added non-negative variable is called a slack variable (or residual variable), and may also be referred to collectively as a slack variable. The coefficients of the newly added slack variable are generally considered to be zero in the objective function.

For the embodiment of the application, in the process of calculating the weight of the content evaluation index, the type of the partial constraint condition is the less type, and the less type includes several cases that the constraint condition is "≦" or "<" or "≧" or ">", and the like. In these cases, it is necessary to convert these constraints into a standard form, so that items that are not differentiable are converted into differentiable items. Specifically, in the embodiment of the present application, the target function is relaxed by using a softmax function.

The Softmax function is a function of the form:

wherein theta is_iAnd x is a column vector and x is,

can be exchanged for a function f with respect to x_i(x) In that respect By means of the softmax function, the range of P (i) can be made to be [0, 1%]In the meantime. In the regression and classification problem, theta is usually the parameter to be found by finding theta that maximizes P (i)_iAs the optimum parameter.

In short, the output values are mapped to the interval of [0,1] through the softmax function, and the summation of the mapped output values is 1, so that when the output value is selected finally, the model parameter with the highest probability (i.e. the output value corresponding to the maximum value) can be selected.

The Gradient Descent (Gradient) algorithm is one of the iterative methods that can be used to solve the least squares problem. Gradient descent is one of the most commonly used methods when solving model parameters of machine learning algorithms, i.e. unconstrained optimization problems. When the minimum value of the loss function is solved, iterative solution can be carried out step by step through a gradient descent method, and the minimized loss function and the model parameter value are obtained. The calculation process of the gradient descent method is to solve the minimum value along the descending direction of the gradient (or solve the maximum value along the ascending direction of the gradient).

The iterative formula of the gradient descent algorithm is as follows:

a_k+1＝a_k+ρ_ks^-(k)… … equation 2

Wherein s is^-(k)Representing the negative direction of the gradient, p_kRepresenting the search step in the gradient direction. The gradient direction can be derived by deriving the function. Is generally trueThe step-size method is determined by a linear search algorithm, i.e. the coordinate of the next point is regarded as a_k+1Then find the function satisfying f (a)_k+1) A of the minimum value of_k+1And (4) finishing.

In general, if the gradient vector is 0, it means that an extreme point is reached, and the magnitude of the gradient is also 0. When the gradient descent algorithm is adopted for optimization solution, the termination condition of the algorithm iteration is that the amplitude of the gradient vector is close to 0, and a very small constant threshold value can be set.

Because the gradient descent algorithm is generally used for solving model parameters when an unconstrained optimization problem is solved, when a final objective function is constructed according to the embodiment of the application, corresponding constraint conditions and initial objective functions are determined according to user groups, and the constraint conditions and the initial objective functions are combined together to form a transition objective function. Thus, the optimization problem with the constraint condition is converted into the optimization problem without the constraint condition. And because the form of the partial constraint condition is an undifferentiable item, the transition objective function is relaxed by utilizing the softmax function, namely the undifferentiable item in the transition objective function is converted into a differentiable item to obtain a final objective function, so that the final objective function is a linear function of the weight, the gradient of the final objective function for the weight can be solved, and the optimal solution of the weight is output by utilizing a gradient descent algorithm.

In the embodiment of the application, the calculation of the estimation value of the content evaluation index and the calculation of the weight of the content evaluation index can be performed off-line; or the calculation of the estimated value of the content evaluation index is executed off-line, and the calculation of the weight of the content evaluation index is executed on-line; or the calculation of the weight of the content evaluation index is executed off-line, and the calculation of the estimation value of the content evaluation index is executed on-line; or the calculation of the estimated value of the content evaluation index and the calculation of the weight of the content evaluation index are both performed online. According to the training samples corresponding to the user groups, after the weight of the content evaluation index of the user group is calculated, the weight of the content evaluation index is stored in association with the user group, so that when the score of the content to be recommended is calculated on line, the weight of the corresponding content evaluation index can be directly obtained. In addition, the recommended content pushed to the user can also be used as a training sample, and the interactive data of the user aiming at the recommended content is acquired and input into the optimization algorithm model for training, so that the weight of the content evaluation index is further optimized.

According to the method and the device, the calculation process of the estimated value of the content evaluation index and the calculation process of the weight are decoupled from the sequencing process of the content to be recommended. After the estimation values of content evaluation indexes (click rate, praise rate, negative feedback rate and the like) of the content to be recommended are obtained through a model fusion mode, an optimization algorithm model is established for a multi-objective optimization problem, the weight of each index prediction value is directly solved, and then the estimation values and the weights are used for calculating scores on line, so that the process of solving the weight through trial and error of a plurality of groups of on-line experiments is avoided, and the calculation difficulty and the calculation amount are reduced.

The method and the device for evaluating the content of the mobile terminal determine the estimation value and the weight of the content evaluation index based on an artificial intelligence technology, specifically calculate the estimation value of the content evaluation index by using a machine learning algorithm model, and calculate the weight of the content evaluation index by using a multi-objective optimization algorithm. In the present application, the model for calculating the estimated value of the content evaluation index is not limited to the deep neural network model, and is not limited to the calculation of the estimated value of the content evaluation index using the machine learning algorithm. On the other hand, the multi-objective optimization algorithm model in the embodiment is only an example and is not limited, and besides the improved gradient descent algorithm, the weight of the content evaluation index can be solved by using algorithms such as a simulated annealing algorithm and a genetic algorithm.

After introducing the design idea of the embodiment of the present application, an application scenario set by the present application is briefly described below. The following scenarios are only used to illustrate the embodiments of the present application and are not limiting. In specific implementation, the technical scheme provided by the embodiment of the application can be flexibly applied according to actual needs.

Fig. 1 is a schematic diagram of a content recommendation system according to an embodiment of the present application. The application scenario includes a terminal device 101, a media host server 102, and an advertisement trading platform server 103, where the terminal device 101, the media host server 102, and the advertisement trading platform server 103 may be directly or indirectly connected through wired or wireless communication, and the application is not limited herein.

The terminal device 101 is configured to send an advertisement request to the advertisement trading platform server 103, and may be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, or a vehicle-mounted terminal, but is not limited thereto. The client corresponding to the media host server 102 is installed in the terminal device 101, and the client may be a web page client, a client installed in the terminal device 101, a light application embedded in a third-party application, or the like. The client embeds an advertisement slot of an advertisement trading platform, and the advertisement trading platform server 103 puts an advertisement of an advertiser on the advertisement slot of the client, thereby displaying the advertisement to a user.

The advertisement transaction platform server 103 may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing basic cloud computing services such as a cloud service, a cloud database, cloud computing, a cloud function, cloud storage, a network service, cloud communication, a middleware service, a domain name service, a security service, a CDN, a big data and artificial intelligence platform, and is applied to an advertisement delivery product to meet a processing requirement of an individualized advertisement delivery amplified data volume.

When implemented based on cloud technology, the advertisement trading platform server 103 may process the user data and the advertisement data in a cloud computing and cloud storage manner.

Cloud computing (cloud computing) is a computing model that distributes computing tasks over a large pool of resources, enabling various application systems to obtain computing power, storage space, and information services as needed. The network that provides the resources is referred to as the "cloud". Resources in the "cloud" appear to the user as being infinitely expandable and available at any time, available on demand, expandable at any time, and paid for on-demand. The cloud computing resource pool mainly comprises: computing devices (which are virtualized machines, including operating systems), storage devices, and network devices.

A distributed cloud storage system (hereinafter, referred to as a storage system) refers to a storage system that integrates a large number of storage devices (storage devices are also referred to as storage nodes) of different types in a network through application software or application interfaces to cooperatively work by using functions such as cluster application, grid technology, and a distributed storage file system, and provides a data storage function and a service access function to the outside.

In a possible implementation mode, user characteristic data, content characteristic data and interaction data of a user for recommended content are stored in a cloud storage mode, when an optimization algorithm model needs to be trained, training samples are obtained from a storage system corresponding to the cloud storage, the optimization algorithm model is trained by the training samples, weights of content evaluation indexes are obtained, at the moment, computing tasks are distributed in a large number of resource pools in a cloud computing mode, computing pressure is reduced, and meanwhile, a training result can be obtained.

The deep neural network model is trained, training samples are obtained from a storage system corresponding to cloud storage, the deep neural network model is trained by the training samples, and when the deep neural network model is trained, computing tasks are distributed in a large number of resource pools in a cloud computing mode, so that computing pressure is reduced, and meanwhile, a training result can be obtained. The trained deep neural network model can be used for predicting an estimated value of a content evaluation index of content to be recommended, when the estimated value of the content evaluation index of the content to be recommended needs to be determined, a content characteristic value and a user characteristic value of the content to be recommended are obtained from a storage system corresponding to cloud storage, the content characteristic value and the user characteristic value are used for prediction, the estimation of the estimated value of the content evaluation index can be performed through a trained machine learning algorithm, at the moment, computing tasks are distributed in a large number of resource pools in a cloud computing mode, computing pressure is reduced, and meanwhile, a prediction result can be obtained.

When the content to be recommended needs to be graded and ranked, the weight and the estimation value of the content evaluation index are obtained from the storage system corresponding to the cloud storage, and the grade of the content to be recommended is calculated by using the weight and the estimation value of the content evaluation index.

The following describes a scenario in which the content recommendation process is applicable.

The advertisement trading platform server 103 trains the deep neural network model based on the collected user characteristic data, advertisement characteristic data and the interaction behavior data of the user for the advertisement in the sample content to obtain corresponding model parameters. And calculating an estimated value of the content evaluation index of the advertisement by using the trained deep neural network model.

The advertisement trading platform server 103 calculates the weight of the content evaluation index offline using the optimization algorithm model. The ad exchange server 103 needs to determine the initial objective function and constraints of the user grouping, and the estimated value of the content rating index of the recommended ad. And calculating an estimated value of the content evaluation index of the recommended advertisement by using the trained deep neural network model. And after the final objective function is formed by utilizing the initial objective function and the constraint condition, the optimal solution of the weight parameter in the final objective function is solved according to a gradient descent algorithm by utilizing the estimated value of the content evaluation index of the recommended advertisement, and the optimal solution is used as the weight of the content evaluation index. The advertisement trading platform server 103 calculates the weight of the content evaluation index corresponding to the user group, and stores the weight of the content evaluation index in association with the user group.

When the user operates on the terminal device, the client sends an advertisement request to the advertisement trading platform server 103 in response to the operation that the user triggers the advertisement request. Here, the terminal device may directly send the advertisement request to the advertisement trading platform server 103, or after the terminal device sends the advertisement request to the media host server 102, the media host server 102 forwards the advertisement request to the advertisement trading platform server 103.

After receiving the advertisement request, the advertisement trading platform server 103 determines the user group of the target user and obtains the weight of the corresponding content evaluation index, determines a plurality of related advertisements matched with the target user, filters the related advertisements to obtain a plurality of advertisements to be recommended, and calculates the pre-estimated value of the content evaluation index of each advertisement to be recommended. Here, the estimated value of the content evaluation index of the advertisement to be recommended is also calculated by the trained deep neural network model. And calculating the score of the advertisement to be recommended by utilizing the weight and the estimated value of the content evaluation index aiming at each advertisement to be recommended. And sequencing the advertisements to be recommended according to the scores, determining the recommended advertisements from the advertisements, and sending the advertisements to the terminal equipment 101.

After the terminal device 101 presents the advertisement to the user, the advertisement trading platform server 103 may use the advertisement as a training sample, collect a content characteristic value of the advertisement, a user characteristic value of a corresponding user, and interaction behavior data of the user for the advertisement, and train and update the deep neural network model and the optimization algorithm model.

It is to be noted that the above-mentioned application scenarios are merely illustrated for the convenience of understanding the spirit and principles of the present application, and the present application embodiments are not limited in any way in this respect. Rather, the embodiments of the present application may be applied to any applicable scenario.

The content recommendation method provided in the embodiment of the present application is described below with reference to the application scenario shown in fig. 1.

Referring to fig. 2, an embodiment of the present application provides a content recommendation method, as shown in fig. 2, the method includes:

step S200: the terminal device transmits a content recommendation request to the server in response to a user operation.

The content recommendation request may be to request to push an advertisement, or may also be to request to push other data or information, such as to push a joke, push a weather forecast, push a service item, and the like.

In the specific operation process, the user operation responded by the terminal device may be an operation actively initiated by the user to request the advertisement, such as clicking an advertisement display area in a webpage, or a special operation preset by the user, such as double-clicking a screen. The method may also be an operation of binding when the user performs other operations, for example, when the user opens a certain webpage, the sending of the advertisement request is triggered, and although the user does not actively initiate the advertisement request, the server still sends the advertisement to the terminal device, so that the terminal device presents the advertisement to the user.

Step S201: the server determines the user group to which the target user belongs.

In the specific implementation process, the users may be grouped according to their own characteristics, such as the liveness of the users, or according to the interaction behavior of the users with the content, such as the interaction frequency of the users with the advertisements, or according to other rules. Because the advertisement trading platform converts the user traffic into cash revenue and also needs to give consideration to the user experience to ensure the good development of ecology, the user grouping is mainly performed according to the historical interaction behavior data of the user and the content in the embodiment of the application.

For example, the number of advertisement exposures and the number of interactions with the advertisement of each user in the past period (e.g. 90 days) are counted, and users with low interaction frequency, common interaction frequency and high interaction frequency with the advertisement are mined according to the exposure threshold and the interaction threshold. Aiming at the users in different user groups, different objective functions and constraint conditions can be adopted to further obtain the weights of different content evaluation indexes. For example, the objective function is set according to the principle of experience priority, or the objective function is set according to the principle of income experience balance, so that the objective functions of different user groups are different.

Step S202: and determining content evaluation index information corresponding to the user group, wherein the content evaluation index information is obtained by calculation according to historical behavior data of the users in the user group aiming at the recommended content.

That is, the content evaluation index information is calculated from the historical behavior data of the users in the user group corresponding to the recommended content. Specifically, the user may perform some operations on the recommended content, such as clicking, evaluating, and the like, and the operations may be recorded in a log or running data of the recommended content. In the embodiment of the application, historical behavior data are used for representing the operations, and content evaluation index information is calculated according to the historical behavior data.

Wherein the content evaluation index information includes at least two content evaluation indexes and a weight of each content evaluation index. Here, the weight corresponding to each content evaluation index is calculated according to sample content obtained from historical behavior data of the recommended content by the users in the user group.

The content evaluation indexes are the basis for grading and sequencing the contents to be recommended, and each content evaluation index corresponds to one item in a grading formula for calculating the grading of the contents to be recommended, such as eCPM, pCTR, the approval rate of the user to the advertisement, the negative feedback rate and the like.

In a specific implementation process, content evaluation index information corresponding to different user groups may be the same or different, where the content evaluation index information is different and may be different, for example, the content evaluation index of the user group a includes eCPM, pCTR and a negative feedback rate, and the internal evaluation index of the user group B includes a click rate, a like rate and a negative feedback rate. For example, the content evaluation index information of the user group a and the content evaluation index information of the user group B are eCPM, click rate and negative feedback rate, but the weight of the user group a is 20%, 40% and 30% in this order, and the weight of the user group B is 50%, 10% and 40% in this order. And the sum of all weights corresponding to the same content to be recommended is 1.

Step S203: and evaluating each content to be recommended of the target user according to the content evaluation index information.

In the specific implementation process, the manner of evaluating the content to be recommended is not limited, for example, all the content to be recommended of the target user may be classified according to the content evaluation index information, and then the recommended content is determined according to the classification result; or labeling each content to be recommended, and determining recommended content according to the label content; or scoring each content to be recommended according to the content evaluation index information, and selecting the recommended content according to the score.

Preferably, in the embodiment of the present application, the evaluation of each content to be recommended is expressed in a scoring manner. The evaluation and recommendation are carried out by using a grading mode, the result is clear, and the method is simple, convenient, easy to operate and easy to operate.

the estimation value of the content evaluation index of the content to be recommended is obtained by estimation according to the content characteristic value of the content to be recommended and the user characteristic value of the target user.

Specifically, the evaluation value of the content evaluation index of each content to be recommended and the corresponding weight may be weighted. Since each user group corresponds to a set of weights, different contents to be recommended of the target user share a set of weights. For example, the content evaluation indexes of the content to be recommended 1, the content to be recommended 2 and the content to be recommended 3 are eCPM, click rate and negative feedback rate, and m is used respectively₁、m₂、m₃It is shown that the corresponding weights are 30%, 20%, and 50%, respectively, and the recommendation scores P of the content to be recommended 1, the content to be recommended 2, and the content to be recommended 3 all satisfy the following formula:

P＝30％m₁+20％m₂+30％m₃… … equation 3

Wherein m is₁eCPM, m for content to be recommended₂For click-through rate of content to be recommended, m₃Is the negative feedback rate of the content to be recommended. The recommendation scores of the content to be recommended 1, the content to be recommended 2 and the content to be recommended 3 are determined by the values corresponding to the content evaluation indexes. In the embodiment of the application, in order to simplify the process and reduce the experimental amount and the data amount, the estimation value is used as the value of the content evaluation index and is substituted into the formula 3, so that the recommendation score of each content to be recommended is calculated. Here, the estimation value is estimated according to the content characteristic value of the corresponding content to be recommended and the user characteristic value of the target userAnd (4) obtaining.

Specifically, the evaluation value of the content evaluation index of the content to be recommended is determined according to the following manner:

inputting the content characteristic value of the content to be recommended and the user characteristic value of the target user into the trained deep neural network model to obtain an estimated value of a content evaluation index of the content to be recommended;

In a specific implementation process, the content evaluation index comprises a user experience index and an electronic resource index associated with the content to be recommended. User experience indexes such as pCTR, like rate of approval, negative feedback rate and the like, and electronic resource indexes associated with the content to be recommended such as CPM, eCPM and the like.

The estimated value of the electronic resource index can be obtained by direct conversion estimation, for example, eCPM reflects the exposure bid of the advertiser for the advertisement, and according to the different advertisement types, the advertiser can directly bid for the exposure, or the advertiser can convert the converted bid into the exposure bid by the platform.

The estimated value of the user experience index can be obtained by utilizing an online experiment, for example, the click rate, a set number of experimental users can be selected, each content to be recommended is pushed to the experimental users, the click rate of the experimental users is counted and used as the estimated value of the click rate, but the method has a long period and is difficult to implement. In the embodiment of the application, the estimation value of the user experience index is obtained by calculation of an algorithm model, and the algorithm model can be a trained deep neural network model. Optionally, the algorithm model may also be a statistical model, or a logistic regression algorithm, etc.

The estimation values in the embodiment of the application comprise the estimation value of the user experience index of the recommended content and the estimation value of the user experience index of the content to be recommended, the two estimation values can be calculated through different algorithm models, but in order to ensure that the estimation standards are unified, the estimation value of the user experience index of the recommended content and the estimation value of the user experience index of the content to be recommended are calculated through the same algorithm model, so that the accuracy can be improved, the calculation steps can be reduced, and the calculation difficulty is reduced.

In addition, the weight of the content evaluation index may be stored after the offline calculation is completed, and may be directly acquired from the storage area when the score is calculated. Of course, the weight of the content evaluation index may also be directly calculated, and is not limited here.

Step S204: and determining recommended content recommended to the target user from all the contents to be recommended according to the evaluation result.

In a specific implementation process, all the contents to be recommended may be ranked according to the scores, and the N contents to be recommended with the best ranking may be used as the recommended contents, or the contents to be recommended with the highest score may be directly used as the recommended contents.

Step S205: the server transmits the recommended content to the terminal device.

Further, before determining the recommended content recommended to the target user from all the contents to be recommended according to the evaluation result, step 204 further includes:

In a specific implementation process, after receiving a content recommendation request of a target user, a server acquires a plurality of related contents matched with the target user from a content storage area or a network. And then filtering through various dimensions, filtering out the related content which does not accord with the filtering rule from all related content, and taking the rest related content as the content to be recommended. The filtering rules include whether the advertisement budget is greater than a budget threshold, whether industry freshness is greater than a freshness threshold, whether eCPM is greater than a revenue threshold, and the like. The preliminary screening can save the calculation amount and improve the recommendation accuracy, and the subsequent grading and sorting processes are all executed in the contents to be recommended.

How to determine the weight corresponding to the content evaluation index is described in detail below.

In the embodiment of the present application, the weight corresponding to the content evaluation index corresponds to the user group, and may be determined according to the following manner:

at least obtaining an estimation value of a content evaluation index of recommended content and a final objective function corresponding to user grouping, wherein the final objective function comprises a weight parameter of the content evaluation index of the recommended content;

determining the gradient of the final objective function aiming at the weight parameter of the content evaluation index;

In the specific implementation process, users in the same user group correspond to the same content evaluation index and the weight of the content evaluation index. The specific content evaluation index and the weight of the content evaluation index are determined according to which content evaluation indexes of the winning content to be recommended need to be optimized. For example, for users with low interaction frequency with advertisements, the interaction behavior frequency of the users, namely click rate, praise rate, forward comment and the like, needs to be increased, and platform benefits can be sacrificed appropriately to achieve better effects. For users with normal advertisement interaction frequency, the platform profit can be optimized on the basis of ensuring the experience indexes. Since platform revenue and user experience indexes are often in a restrictive relationship, the embodiments of the present application need to explicitly optimize targets and quantify the balance between the targets through a modeling manner.

Suppose that under a certain user group, the total number of times of requesting advertisements by users in unit time is K, that is, the users have K queues of advertisements to be recommended in total. Determining the estimation values of content evaluation indexes such as eCPM, pCTR, like rate of praise, negative feedback rate and the like for each advertisement j to be recommended in the queue k, and finally obtaining the recommendation score S_k,j. After the advertisements in the queue k are sorted by the score, the advertisement i is recommended_kFor the highest scoring ad, namely:

i_k＝argmax_j{S_k,j… … equation 4

Wherein S is_k,jThe recommendation score of the ad j to be recommended in the queue k.

The modeling process is illustrated with a target user having a normal frequency of interaction with the advertisement. For the part of target users, the optimization aims to maximize platform benefits on the basis of ensuring user experience indexes. Since revenue generated by a single pull is measured by eCPM, the objective function is modeled as an average eCPM that optimizes the recommended advertisements, formulated as follows:

wherein ecpm_ikThe eCPM of advertisement i to be recommended in queue k.

In addition, a baseline of user experience metrics under the user group and a tolerance range of advertising category fluctuation need to be established. For click-through rate pCTR, the baseline for the average pCTR of the recommended ads is pCTR₀That is, the average pCTR of all recommended advertisements needs to be greater than pCTR₀Then the corresponding constraint is:

wherein, pctr_ikThe click rate of the advertisement i to be recommended in the queue k.

Similarly, other user experience indicators similar to equation 6 may be added, such as the like, the like. In addition, let

Representing categories that need to control fluctuations (e.g., focus-type advertisements, download-type advertisements, etc.), by

Representing that the advertisement to be recommended belongs to the category (1 represents that the advertisement to be recommended belongs to the category, and 0 represents that the advertisement to be recommended does not belong to the category), the upper limit and the lower limit of the category ratio are r respectively_uAnd r_lThen the constraint for category fluctuation is:

wherein the content of the first and second substances,

for the purpose of the category in which the fluctuations need to be controlled,

belong to for the advertisement to be recommended in the queue k

Class (iii) of the subject.

And (4) combining the formulas 4 to 8, so that the optimization problem under the user grouping can be modeled. Wherein the recommendation score S_k,jDecides to recommend advertisement i_kAnd further determines the values of the objective function and the constraint condition. According to the multi-objective optimization problem, corresponding recommendation score S is calculated_k,jThe contents evaluation indexes eCPM, pCTR,

(representing whether advertisement j belongs to a category

1 if yes, 0 if no). Recommendation score S_k,jThe formula of (1) is as follows:

wherein, ω is₁、ω₂、ω₃As a weight of content evaluation index, ecpm_k,jeCPM, pctr for advertisement j to be recommended in queue k_ikFor the click rate of the advertisement j to be recommended in the queue k,

belong to for the advertisement to be recommended in the queue k

Class (iii) of the subject. When constraint conditions such as a praise rate threshold and other category fluctuation thresholds are added to the optimization problem, corresponding content evaluation index items and weights are added to the formula 9.

Assuming that there is an objective function, m user experience index constraints, and n category fluctuation constraints in the optimization problem, the formula 9 correspondingly contains m + n +1 content evaluation index items, which also correspond to m + n +1 weights. In the embodiment of the application, the optimal solution of m + n +1 weights ω in the formula 9 is solved through offline training, so that the value of the formula 5 is the maximum under the condition of satisfying the constraint conditions of the formulas 6 to 8.

Similarly, the objective function and constraints of the multi-objective optimization problem may be changed for users who have a low frequency of interaction with the advertisement. For example, the average praise number of the user may be used as an objective function, and the average eCPM may be used as a constraint condition. Similarly, the optimal solution of m + n +1 weights ω in equation 9 needs to be solved, so that the average pCTR of the recommended advertisement is the maximum when the constraint condition is satisfied.

In the embodiment of the invention, after the optimization problem under different user groups is modeled, a unified optimization algorithm model can be adopted to solve. Fig. 3 shows a schematic structural diagram of an optimization algorithm model in the embodiment of the present application.

As shown in FIG. 3, the data used to train the multi-objective optimization algorithm model includes initial objective functions and constraints corresponding to the user groups. In the embodiment of the application, multiple optimization targets need to be achieved simultaneously through a multi-objective optimization algorithm, one most main optimization target can be determined, a corresponding initial objective function is formed, and the rest optimization targets are used as constraint conditions. In the example shown in fig. 3, the most important optimization objective is to optimize platform revenue or user experience, and therefore, an initial objective function is formed and input into the multi-objective optimization algorithm based on the optimized platform revenue or user experience. The other optimization objectives may include optimizing user experience, optimizing platform revenue, and optimizing category fluctuation, such that one or more constraints, such as user experience constraints, platform revenue constraints, and category fluctuation constraints, may be formed according to the optimization objectives, and a multi-objective optimization algorithm may be input to achieve the objective of balancing various optimization requirements. Further, after training of the multi-objective optimization algorithm, content evaluation indexes for scoring the content to be recommended and corresponding weights can be calculated. The process of deriving weights by multi-objective optimization algorithm training is described in detail below.

The offline data includes:

the user side: and grouping users, wherein the baselines of the objective function and the constraint condition under the corresponding user grouping are corresponding to the user grouping. For example, for a user group with normal frequency of interaction with the advertisement, the optimization goal is to maximize the average eCPM, and the baseline index of the constraint is pctr₀、r_uAnd r_lAnd the like. Aiming at the user groups with low interaction frequency with the advertisements, the optimization target is to maximize the average pCTR, and the baseline index of the constraint condition is ecpm₀、r_uAnd r_lAnd the like.

And (3) advertisement side: an estimate of a content rating indicator for each recommended advertisement in the queue of all recommended advertisements, e.g., ecpm for each advertisement j in queue k_k,j，pctr_k,j,

And the like. Wherein the evaluation value of the content evaluation index of the recommended content is determined according to the following mode: and inputting the content characteristic value of the recommended content and the user characteristic value of the recommended content into the trained deep neural network model, and calculating to obtain an estimation value of the content evaluation index of the recommended content.

The embodiment of the application collects all advertisement data in the advertisement queue corresponding to each group of user groups. And under the corresponding user grouping, solving the weight according to the multi-objective optimization problem modeled by the user grouping and the off-line data, and finally outputting each content evaluation index and the corresponding weight in the formula 9.

In a specific implementation process, the weights can be solved by using a simulated annealing algorithm, a genetic algorithm and the like. In the embodiment of the application, the relaxation problem is combined with a gradient descent algorithm, specifically, a target function and a constraint condition are combined together in a form of a penalty term, so that the gradient descent algorithm can be used for calculation, and further, an item which is not differentiable in a calculation formula is converted into a differentiable item through a softmax function, so that the calculation process is simplified on the basis of ensuring that the weight can be calculated.

Correspondingly, the final objective function corresponding to the user group in the embodiment of the present application is determined according to the following manner:

determining a constraint condition and an initial objective function corresponding to a user group;

and determining a final objective function according to the constraint conditions and the initial objective function.

Determining a final objective function according to the constraint condition and the initial objective function, wherein the determining the final objective function according to the constraint condition and the initial objective function comprises the following steps:

combining the constraint condition and the initial objective function into a transition objective function;

In the specific implementation process, firstly, an initial objective function and constraint conditions of user grouping are determined, and the constrained optimization problem is converted into an unconstrained optimization problem. The constraint conditions plus the relu function and the coefficient λ, and the initial objective function (i.e. equation 4) are combined into a transition objective function:

relaxation is performed using the softmax function, which transforms the non-differentiable terms in equations 7 and 8 into differentiable terms:

substituting equation 11 into equation 10 results in the objective function f (S) in equation 10_k,j) For S_k,jA gradient exists. Also, as can be seen from equation 9, S_k,jIs a linear function of the weight w. Thus, by the chain rule, the objective function f (S) can be solved_k,j) Gradient for w. To obtainAfter the numerical solution of the gradient, iterative optimization is carried out by using a gradient descent method, and an optimal solution { w is output^＊And f, the weight is obtained.

The content evaluation indexes corresponding to the formulas 4-8 are eCPM item, pCTR item, and category factor obtained by the multi-objective optimization algorithm model shown in FIG. 3

The corresponding weight is sequentially

Namely, the calculation formula of the recommendation score is as follows:

thus, the weight { w of the content evaluation index output by the multi-objective optimization algorithm model can be utilized^＊And calculating the recommendation score S of each advertisement to be recommended according to the estimated value of the content evaluation index_k,jAnd sequencing the advertisements to be recommended, and finally determining the recommended advertisements.

The above flow is described in detail below with specific embodiments, and the specific flow of the specific embodiments is shown in fig. 4, and includes:

the server receives an advertisement recommendation request sent by the terminal equipment, wherein the advertisement recommendation request comprises a user identifier of a target user.

The server provides an online advertisement recommendation service to the target user based on the advertisement recommendation request. Firstly, the server determines the user group of the target user, and determines the ranking factors corresponding to the user group and the weight of each ranking factor. The ranking factor is the content evaluation index in the above embodiment. And the server determines all the matched related advertisements of the target user and filters all the related advertisements to obtain a plurality of advertisements to be recommended. And the server estimates the ranking factor estimation value of each advertisement to be recommended by using the deep neural network model, and acquires the ranking factors and the weights corresponding to the user groups. And the server performs weighted calculation according to the ranking factor estimated value and the ranking factor weight to obtain the score of each advertisement to be recommended, and ranks all the advertisements to be recommended according to the scores. And according to the sequence, the server determines the recommended advertisement sent to the target user and sends the recommended advertisement to the user.

The estimated value of the ranking factor in the above embodiment is pre-estimated by a deep neural network model, and the ranking factor and the weight are calculated by a multi-objective optimization algorithm. The process of calculating the weight of the ranking factor is performed off line, after the server determines the ranking factor and the corresponding weight, the weight of the ranking factor and the user group can be stored in a related mode, and when the advertisements to be recommended need to be ranked, the advertisements can be directly inquired and obtained.

In addition, the server can take the advertisements recommended to the user as training samples, collect user interaction behavior data of the user aiming at the recommended advertisements, and train and update the deep neural network model and the multi-objective optimization algorithm.

The following are embodiments of the apparatus of the present application, and for details not described in detail in the embodiments of the apparatus, reference may be made to the above-mentioned one-to-one corresponding method embodiments.

Referring to fig. 5, a block diagram of a content recommendation apparatus according to an embodiment of the present application is shown. The cross-link data processing apparatus is implemented by hardware or a combination of hardware and software as all or a part of the server 103 in fig. 1. The device includes: grouping unit 501, index unit 502, evaluation unit 503, determination unit 504, and filtering unit 505.

the index unit is used for determining content evaluation index information corresponding to the user group, and the content evaluation index information is obtained by calculation according to historical behavior data of the users in the user group aiming at recommended content;

In an alternative embodiment, the content evaluation index information includes at least two content evaluation indexes and a weight of each content evaluation index;

the weight corresponding to each content evaluation index is calculated according to sample content obtained by the users in the user group aiming at the historical behavior data of the recommended content.

In an optional embodiment, the evaluation unit is specifically configured to determine the evaluation value of the content evaluation index of the content to be recommended according to the following manner:

inputting the content characteristic value of the recommended content and the user characteristic value of the recommended content into the trained deep neural network model, and calculating to obtain an estimated value of the content evaluation index of the recommended content;

and converting the incautizable item in the transition objective function into a incautizable item to obtain a final objective function.

In an optional embodiment, the apparatus further comprises a filter unit for:

Referring to fig. 6, a block diagram of a server according to an embodiment of the present application is shown. The server 1100 is implemented as the server 103 in fig. 1. Specifically, the method comprises the following steps:

the server 1100 includes a Central Processing Unit (CPU)801, a system memory 1104 including a Random Access Memory (RAM)1102 and a Read Only Memory (ROM)1103, and a system bus 1105 connecting the system memory 1104 and the central processing unit 1101. The server 1100 also includes a basic input/output system (I/O system) 1106, which facilitates transfer of information between devices within the computer, and a mass storage device 1107 for storing an operating system 1113, application programs 1114, and other program modules 1115.

The basic input/output system 1106 includes a display 1108 for displaying information and an input device 1109 such as a mouse, keyboard, etc. for user input of information. Wherein the display 1108 and the input device 1109 are connected to the central processing unit 1101 through an input output controller 1110 connected to the system bus 1105. The basic input/output system 1106 may also include an input/output controller 1110 for receiving and processing input from a number of other devices, such as a keyboard, mouse, or electronic stylus. Similarly, input-output controller 1110 also provides output to a display screen, a printer, or other type of output device.

The mass storage device 1107 is connected to the central processing unit 1101 through a mass storage controller (not shown) that is connected to the system bus 1105. The mass storage device 1107 and its associated computer-readable media provide non-volatile storage for the server 1100. That is, the mass storage device 1107 may include a computer-readable medium (not shown) such as a hard disk or CD-ROM drive.

Without loss of generality, computer readable media may comprise computer storage media and communication media. Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media includes RAM, ROM, EPROM, EEPROM, flash memory or other solid state memory technology, CD-ROM, DVD, or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices. Of course, those skilled in the art will appreciate that computer storage media is not limited to the foregoing. The system memory 1104 and mass storage device 1107 described above may be collectively referred to as memory.

The server 1100 may also operate in accordance with various embodiments of the application through remote computers connected to a network, such as the internet. That is, the server 1100 may connect to the network 1112 through the network interface unit 1111 that is coupled to the system bus 1105, or may connect to other types of networks or remote computer systems (not shown) using the network interface unit 1111.

The memory also includes one or more programs, stored in the memory, the one or more programs including instructions for performing the content recommendations provided by embodiments of the present application.

It will be understood by those skilled in the art that all or part of the steps in the content recommendation method of the above embodiments may be implemented by a program to instruct associated hardware, where the program may be stored in a computer-readable storage medium, where the storage medium may include: read Only Memory (ROM), Random Access Memory (RAM), magnetic or optical disks, and the like.

It will be understood by those skilled in the art that all or part of the steps in the content recommendation method of the above embodiments may be implemented by a program to instruct associated hardware, where the program may be stored in a computer-readable storage medium, and the storage medium may include: read Only Memory (ROM), Random Access Memory (RAM), magnetic or optical disks, and the like.

The above-mentioned serial numbers of the embodiments of the present application are merely for description and do not represent the merits of the embodiments.

The above description is only exemplary of the present application and should not be taken as limiting the present application, as any modification, equivalent replacement, or improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A content recommendation method, comprising:

determining a user group to which a target user belongs;

determining content evaluation index information corresponding to the user group, wherein the content evaluation index information is obtained by calculation according to historical behavior data of users in the user group aiming at recommended content;

and determining recommended content recommended to the target user from all the contents to be recommended according to the evaluation result.

2. The method according to claim 1, wherein the content evaluation index information includes at least two content evaluation indexes and a weight of each content evaluation index;

3. The method according to claim 2, wherein the evaluating each content to be recommended of the target user according to the content evaluation index information comprises:

4. The method according to claim 3, wherein the evaluation value of the content evaluation index of the content to be recommended is determined according to the following manner:

5. The method according to claim 2, wherein the weight corresponding to the content evaluation index is determined according to the following manner:

6. The method according to claim 5, wherein the evaluation value of the content evaluation index of the recommended content is determined according to the following manner:

7. The method of claim 5, wherein the final objective function for the user group is determined according to the following:

8. The method of claim 7, wherein determining the final objective function based on the constraints and the initial objective function comprises:

9. The method according to claim 1, wherein before determining the recommended content recommended to the target user from all the contents to be recommended according to the evaluation result, the method further comprises:

10. A content recommendation apparatus characterized by comprising:

11. The apparatus according to claim 10, wherein the content evaluation index information includes at least two content evaluation indexes and a weight of each content evaluation index;

12. The device according to claim 11, characterized in that the evaluation unit is specifically configured to:

13. The apparatus according to claim 12, wherein the evaluation unit is specifically configured to determine the weight corresponding to the content evaluation index according to the following manner:

14. A computer device, comprising:

at least one processor, and

a memory communicatively coupled to the at least one processor;

wherein the memory stores instructions executable by the at least one processor, the at least one processor implementing the method of any one of claims 1-9 by executing the instructions stored by the memory.

15. A computer-readable storage medium having stored thereon computer instructions which, when executed on a computer, cause the computer to perform the method of any one of claims 1 to 9.