An information recommendation method and device, and a storage medium

Abstract

The application discloses a method and device for information recommendation based on a machine algorithm, and the method can also be applied to the field of big data. The application comprises the following steps: acquiring a user portrait feature corresponding to a candidate user of to-be-recommended information; determining an exposure click feature corresponding to the candidate user according to an information recommendation state of the to-be-recommended information of the candidate user in a historical time period; determining an estimated exposure rate, an estimated click rate and an estimated conversion rate of the to-be-recommended information for the candidate user according to the user portrait feature and the exposure click feature of the candidate user; determining an estimated delivery result of the to-be-recommended information for the candidate user according to the estimated exposure rate, the estimated click rate and the estimated conversion rate of the to-be-recommended information for the candidate user; and if the estimated delivery result meets an information delivery condition, recommending the to-be-recommended information to a terminal device used by the candidate user, thereby improving the conversion rate of the to-be-recommended information and reducing the information delivery cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of big data, and in particular to a method and device for information recommendation and a storage medium. BACKGROUND

[0002] With the continuous development of social networks, more and more advertisers will choose to place advertisements on social networks. The social network can take the large user group as the target audience to achieve the purpose of advertisement promotion and recommendation. Specifically, the advertiser can circle the candidate group in the large user group, and the social network operator pushes the advertisement to the candidate group according to the demand of the advertiser. In this way, the user can see the advertisement recommendation when using the social network, so as to purchase the related product according to the advertisement recommendation.

[0003] When the advertiser circles the candidate group, the social network operator will select part of the group to expose the advertisement according to the user activity and other conditions. At the same time, part of the users in the group exposed to the advertisement will not open the advertisement link, and the users who open the advertisement link will not necessarily purchase the product, which will greatly affect the promotion effect of the advertisement. Therefore, the social network operator analyzes the large user group to determine the user group interested in the advertisement and recommends it to the advertiser, which will reduce the cost of the advertiser and improve the efficiency of the advertisement recommendation.

[0004] The existing social network operator only analyzes the conversion group that purchases the related product, and simply filters the large user group according to the characteristics of the conversion group, which will result in that the determined targeting group still cannot realize accurate placement, the actual conversion efficiency is very low, the cost of the advertiser is increased, and thus the advertisement placement effect of the social network operator is affected. SUMMARY

[0005] The embodiments of the present application provide a method and device for information recommendation and a storage medium, which can predict the placement result of the to-be-recommended information for the candidate user according to the exposure rate, click rate and conversion rate of the to-be-recommended information for the candidate user, so as to select the final user to recommend the information according to the estimated placement result of the large user group, improve the placement effect of the to-be-recommended information, and reduce the cost of information recommendation.

[0006] Therefore, the present application provides a method for information recommendation, comprising:

[0007] Obtaining the user portrait features corresponding to the candidate user of the to-be-recommended information.

[0008] According to the information recommendation state of the recommended information of the candidate user in the historical time period, the exposure and click features corresponding to the candidate user are determined, wherein the recommended information in the historical time period has an associated relationship with the to-be-recommended information.

[0009] The estimated exposure rate, the estimated click rate, and the estimated conversion rate of the to-be-recommended information for the candidate user are determined according to the user portrait feature and the exposure click feature of the candidate user.

[0010] The estimated delivery result of the to-be-recommended information for the candidate user is determined according to the estimated exposure rate, the estimated click rate, and the estimated conversion rate of the to-be-recommended information for the candidate user.

[0011] If the estimated delivery result meets the information delivery condition, the to-be-recommended information is recommended to the terminal device used by the candidate user.

[0012] Another aspect of the present application provides an information recommendation device, which comprises:

[0013] An acquisition unit is configured to acquire a user portrait feature corresponding to a candidate user of to-be-recommended information.

[0014] A determination unit is configured to determine an exposure click feature corresponding to the candidate user according to an information recommendation state of recommended information of the candidate user in a historical time period, wherein the recommended information in the historical time period has an association relationship with the to-be-recommended information.

[0015] The determination unit is further configured to determine the estimated exposure rate, the estimated click rate, and the estimated conversion rate of the to-be-recommended information for the candidate user according to the user portrait feature and the exposure click feature of the candidate user.

[0016] A processing unit is configured to determine the estimated delivery result of the to-be-recommended information for the candidate user according to the estimated exposure rate, the estimated click rate, and the estimated conversion rate of the to-be-recommended information for the candidate user.

[0017] A sending unit is configured to recommend the to-be-recommended information to the terminal device used by the candidate user if the estimated delivery result meets the information delivery condition.

[0018] In a possible design, the determination unit is specifically configured to input the user portrait feature and the exposure click feature of the candidate user into a first model, determine the estimated exposure rate of the to-be-recommended information for the candidate user through the first model, input the user portrait feature and the exposure click feature of the candidate user into a second model, determine the click rate of the to-be-recommended information for the candidate user through the second model, and input the user portrait feature of the candidate user into a third model, and determine the estimated conversion rate of the to-be-recommended information for the candidate user through the third model.

[0019] In a possible design, the information recommendation device further comprises a training unit, and wherein:

[0020] The acquisition unit is further configured to acquire population targeting data and exposure population data corresponding to at least one historical recommended information, wherein the information type of the historical recommended information is the same as that of the to-be-recommended information.

[0021] The determining unit is further configured to determine unexposed crowd data corresponding to the historical recommendation information according to the crowd orientation data and the exposed crowd data. The exposed crowd data is determined as first positive sample training data, and the unexposed crowd data is determined as first negative sample training data. Each of the first positive sample training data and the first negative sample training data includes user portrait features and exposed click features of a historical user, and an actual exposure result of the historical recommendation information to the historical user.

[0022] The training unit is configured to train a first model according to the first positive sample training data and the first negative sample training data.

[0023] In a possible design, the information recommendation apparatus further includes a training unit, and wherein:

[0024] The obtaining unit is further configured to obtain exposed crowd data and clicked crowd data corresponding to at least one historical recommendation information, wherein the historical recommendation information and the to-be-recommended information are of the same information type.

[0025] The determining unit is further configured to determine exposed non-clicked crowd data corresponding to the historical recommendation information according to the exposed crowd data and the clicked crowd data. The clicked crowd data is determined as second positive sample training data, and the exposed non-clicked crowd data is determined as second negative sample training data. Each of the second positive sample training data and the second negative sample training data includes user portrait features and exposed click features of a historical user, and an actual click result of the historical recommendation information to the historical user.

[0026] The training unit is configured to train a second model according to the second positive sample training data and the second negative sample training data.

[0027] In a possible design, the information recommendation apparatus further includes a training unit, and wherein:

[0028] The obtaining unit is configured to obtain clicked crowd data and converted crowd data corresponding to at least one historical recommendation information, wherein the historical recommendation information and the to-be-recommended information are of the same information type.

[0029] The determining unit is configured to determine clicked non-converted crowd data corresponding to the historical recommendation information according to the clicked crowd data and the converted crowd data. The converted crowd data is determined as third positive sample training data, and the clicked non-converted crowd data is determined as third negative sample training data. Each of the third positive sample training data and the third negative sample training data includes user portrait features of a historical user, and an actual conversion result of the historical recommendation information to the historical user.

[0030] The training unit is configured to train a third model according to the third positive sample training data and the third negative sample training data.

[0031] In a possible design, the determining unit is specifically configured to input the estimated exposure rate, the estimated click rate, and the estimated conversion rate of the to-be-recommended information for the candidate user into the information recommendation estimation model, and output the estimated delivery result of the to-be-recommended information for the candidate user through the information recommendation estimation model.

[0032] In a possible design, the information recommendation apparatus further includes a training unit, and wherein:

[0033] The obtaining unit is further configured to obtain population orientation data corresponding to at least one historical recommended information, the historical recommended information and the to-be-recommended information being of the same information type, the population orientation data including user portrait features and exposure click features of a historical user, and an actual delivery result of the historical recommended information for the historical user.

[0034] The obtaining unit is further configured to input the user portrait features and the exposure click features into a first model to obtain, through the first model, an estimated exposure rate of the historical recommended information for the historical user, input the user portrait features and the exposure click features into a second model to obtain, through the second model, an estimated click rate of the historical recommended information for the historical user, and input the user portrait features into a third model to obtain, through the third model, an estimated conversion rate of the historical recommended information for the historical user.

[0035] The determining unit is further configured to input the estimated exposure rate, the estimated click rate, and the estimated conversion rate of the historical recommended information for the historical user into the information recommendation estimation model, and obtain, through the information recommendation estimation model, the estimated delivery result of the historical recommended information for the historical user.

[0036] The training model is configured to train the information recommendation estimation model according to a loss function, wherein the loss function is used to evaluate the similarity between the actual delivery result and the predicted delivery result of the historical recommended information for the historical user.

[0037] In a possible design, the determining unit is specifically configured to classify the candidate user through the information recommendation estimation model.

[0038] The sending unit is specifically configured to recommend the to-be-recommended information to a terminal device used by the candidate user if the candidate user belongs to the target classified population.

[0039] In a possible design, the determining unit is specifically configured to determine, through the information recommendation estimation model, a delivery score of the to-be-recommended information for the candidate user.

[0040] The sending unit is specifically configured to recommend the to-be-recommended information to a terminal device used by the candidate user if the delivery score exceeds a preset threshold.

[0041] In a possible design, the acquisition unit is further configured to acquire a first information recommendation state of the candidate user in a historical time period with respect to a target information push position, the first information recommendation state including a push state of at least one recommended information on the target information push position, and the target information push position being an information push position corresponding to the to-be-recommended information.

[0042] The determination unit is specifically configured to determine, according to the first information recommendation state, an information exposure number and an information click number of the at least one recommended information on the target information push position, and determine the information exposure number and the information click number of the candidate user on the target information push position as the exposure-click feature corresponding to the candidate user.

[0043] In a possible design, the acquisition unit is further configured to acquire a second information recommendation state of the candidate user in the historical time period with respect to a target recommended information, the second information recommendation state including a push state of the target recommended information on at least one information push position, and the target recommended information being of the same type as the to-be-recommended information.

[0044] The determination unit is specifically configured to determine, according to the second information recommendation state, an information exposure number and an information click number of the target recommended information on the at least one information push position, and determine the information exposure number and the information click number of the target recommended information on the at least one information push position as the exposure-click feature corresponding to the candidate user.

[0045] In a possible design, the acquisition unit is further configured to acquire state information corresponding to the to-be-recommended information, the state information including industry information corresponding to the to-be-recommended information, target push position information, and image-text content information.

[0046] The acquisition unit is further configured to acquire candidate information according to the state information corresponding to the to-be-recommended information, wherein the candidate information is of the same industry as the industry corresponding to the to-be-recommended information, and the candidate information is of the same target push position as the target push position corresponding to the to-be-recommended information.

[0047] The determination unit is further configured to acquire a first feature vector corresponding to the to-be-recommended information according to the image-text content information corresponding to the to-be-recommended information, acquire a second feature vector corresponding to the candidate information according to the image-text content information corresponding to the candidate information, and determine a historical recommended information corresponding to the to-be-recommended information from the candidate information according to a cosine similarity of the first feature vector and the second feature vector.

[0048] Another aspect of the present application provides a computer readable storage medium, which stores instructions, when the instructions are run on a computer, causing the computer to perform the method in the above aspects.

[0049] Another aspect of the present application provides a computer program product or computer program comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the method described in the above aspects.

[0050] From the above technical solutions, the embodiments of the present application have the following advantages:

[0051] In the embodiments of the present application, a method for information recommendation is provided. First, the exposure rate, the click rate and the conversion rate of the to-be-recommended information for the candidate user are estimated according to the user portrait features and the exposure click features of the candidate user, and then the final delivery result of the to-be-recommended information for the user is predicted by analyzing the estimated exposure rate, the estimated click rate and the estimated conversion rate. In this way, the user who is most suitable for delivering the to-be-recommended information can be selected according to the predicted delivery result to deliver the information, so as to achieve the effect of accurately delivering the to-be-recommended information. In this way, the useless information delivery behavior can be reduced, so as to reduce a large amount of redundant information in the Internet, save network resources, improve the conversion rate of the to-be-recommended information, and reduce the information delivery cost. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 A scene diagram of advertisement pushing provided by the embodiments of the present application is provided.

[0053] Figure 2 An environment diagram of an information recommendation system provided by the embodiments of the present application is provided.

[0054] Figure 3 A flow diagram of a first model training method provided by the embodiments of the present application is provided.

[0055] Figure 4A A page display diagram of an advertisement pushing position provided by a social platform provided by the embodiments of the present application is provided.

[0056] Figure 4B Another page display diagram of an advertisement pushing position provided by a social platform provided by the embodiments of the present application is provided.

[0057] Figure 4C Another page display diagram of an advertisement pushing position provided by a social platform provided by the embodiments of the present application is provided.

[0058] Figure 5 A flow diagram of a second model training method provided by the embodiments of the present application is provided.

[0059] Figure 6 A flow diagram of a third model training method provided by the embodiments of the present application is provided.

[0060] Figure 7 A flowchart of a method for training an information recommendation estimation model is provided for an embodiment of the present application;

[0061] Figure 8 A structural diagram of an information recommendation system is provided for an embodiment of the present application;

[0062] Figure 9 A flowchart of a method for training an information recommendation estimation model is provided for an embodiment of the present application;

[0063] Figure 10A A page display diagram of a smart directional request is provided for an embodiment of the present application;

[0064] Figure 10B A page display diagram of a smart directional request is provided for an embodiment of the present application;

[0065] Figure 11 A structural diagram of an information recommendation device is provided for an embodiment of the present application;

[0066] Figure 12 A structural diagram of a server is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0067] The method, device and storage medium for information recommendation provided by the embodiments of the present application can predict the delivery result of the to-be-recommended information for the candidate user according to the exposure rate, click rate and conversion rate of the to-be-recommended information for the candidate user, so that the final user can be selected according to the estimated delivery result of the big board user to recommend the information, the delivery effect of the to-be-recommended information is improved, and the information recommendation cost is reduced.

[0068] The terms "first", "second", "third", "fourth" and the like (if any) in the description and claims of the present application and in the above description of the drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a particular order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented, for example, in an order other than that illustrated or described herein. In addition, the terms "include" and "correspond to" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0069] Nowadays, social platforms attract a large number of users as a platform for sharing opinions, insights and experiences with each other, and the information spread on the social platform has become the most important content for users to browse the Internet. As the most powerful information dissemination media, the social platform not only can meet the needs of users to communicate with each other, but also provides a platform for various information push such as advertisement push, video recommendation, text information push, etc. The following will take the advertisement push as an example to introduce the scheme, the advertiser can publish the advertisement to be recommended to the social platform, and recommend the advertisement to a large number of registered users on the social platform, so as to achieve the purpose of advertisement promotion.

[0070] Compared with traditional advertisement promotion media, the advertisement promotion on the social platform has the characteristics of fast speed, high timeliness, wide spread, etc., so it will be welcomed by advertisers. At the same time, the social platform providing advertisement position can also realize the function of intelligent promotion, that is, to select users and push different types of advertisements to different users, so as to save network resources and reduce the cost of advertisement promotion under the premise of ensuring the effect of advertisement promotion. Therefore, for a specific advertisement, reasonable people orientation becomes the key to advertisement promotion.

[0071] People orientation refers to that before putting the advertisement, the advertiser circles the candidate people in the large customer of the social platform as the orientation condition, that is, puts the advertisement to the users in the candidate people. Figure 1 A scene schematic diagram of advertisement push is provided for the application embodiment. Specifically, when the advertiser plans to put the advertisement on a certain social platform, the candidate people need to be selected first, and the social platform is instructed to put the advertisement to the candidate people. For example, the advertiser can inform the social platform of the users to which the advertisement needs to be put through the form of directional number package. The directional number package can include the identification of each user in the candidate people selected by the advertiser, for example, the directional number package can be a set of social platform accounts of multiple users. When the social platform receives the directional number package, it locks the users to be put according to the account set in the directional number package, and then puts the advertisement to the terminal device used by these users.

[0072] Because the social platform stores a large amount of user data and can also obtain the use behavior of the user to the social platform, the social platform can predict the attitude of the user to a certain advertisement and the putting effect of the advertisement by analyzing the user data and the user behavior. Therefore, the social platform can recommend one or more candidate people to the advertiser for selection, and the social platform can recommend the people who can accept the advertisement as the orientation condition to the advertiser according to the predicted effect, so that the advertiser can obtain more effective users without changing the size of the directional number package, and improve the conversion rate of the advertisement. Therefore, how the social platform analyzes the large plate users, predicts the users who can accept a certain advertisement and recommends them to the advertiser becomes a problem to be solved.

[0073] The method provided in the embodiments of the present application can be implemented based on big data. Big data refers to a collection of data that cannot be captured, managed, and processed by conventional software tools within a certain time frame. It is a massive, high-growth, and diversified information asset that requires new processing models to have stronger decision-making power, insight discovery, and process optimization capabilities. With the advent of the cloud era, big data has also attracted more and more attention. Big data requires special technologies to effectively process large amounts of data within a tolerable time frame. Technologies suitable for big data include large-scale parallel processing databases, data mining, distributed file systems, distributed databases, cloud computing platforms, the Internet, and scalable storage systems.

[0074] Before introducing the technical solutions of the embodiments of the present application, several concepts are explained first:

[0075] 1. Ad exposure rate: Ad exposure rate is also called reach rate. Once an advertiser identifies a candidate group (targeted audience), they will inform the social platform of the candidate group. The social platform needs to analyze each user in the candidate group to determine whether they meet the exposure conditions. For example, it can determine whether the user's browsing time on the social platform exceeds a threshold within a certain period of time, or determine whether the user meets other targeting conditions, such as whether the user's age is within a specified age range. Once a user is determined to meet the exposure conditions, an ad is pushed to that user. Therefore, the ad exposure rate refers to the ratio of the number of users in the candidate group who are exposed to the ad to the total number of users in the candidate group.

[0076] 2. Click-through rate of advertisements: When an advertisement is exposed to users, they have different attitudes towards the advertisements. Some users will simply ignore the advertisements and choose not to view the advertisement content; while some users will click to read or directly browse the advertisements. Therefore, the click-through rate of an advertisement refers to the ratio of the number of users who read the advertisement to the number of users who were exposed to the advertisement.

[0077] 3. Ad conversion rate: When users click and read an ad, some will choose to close the ad, while others will follow the link in the ad to jump to the corresponding shopping page and complete the purchase of the relevant product. Therefore, the conversion rate of an ad refers to the ratio of the number of users who purchase the product to the number of users who click on the ad.

[0078] In combination with the above description, this application proposes a method for information recommendation, which is applied to Figure 2 In the information recommendation system shown in Figure 2 , Figure 2An environment schematic diagram of an information recommendation system in an embodiment of the present application is shown in the figure. The information recommendation system includes a server and a terminal device. The server side completes analysis and prediction of large-dish users on a social platform, obtains a candidate group for information, and recommends the candidate group to an advertiser. The client side implements push display of the information.

[0079] It should be noted that, Figure 2 The server in the information recommendation system can be a server or a server cluster composed of multiple servers or a cloud computing center, and the specific implementation is not limited herein. The client is deployed with a terminal device, which can be a tablet computer, a notebook computer, a palm computer, a mobile phone, a personal computer (PC), and a voice interaction device shown in the figure. Figure 2

[0080] The terminal device and the server can communicate through a wireless network, a wired network, or a removable storage medium. The wireless network uses standard communication technology and / or protocols. The wireless network is usually the Internet, but can also be any network, including but not limited to Bluetooth, a local area network (LAN), a metropolitan area network (MAN), a wide area network (WAN), mobile, a dedicated network, or any combination of a virtual private network. In some embodiments, custom or dedicated data communication technology can be used instead of or in addition to the above data communication technology. The removable storage medium can be a universal serial bus (USB) flash disk, a mobile hard disk, or other removable storage media.

[0081] Although Figure 2 only five terminal devices and one server are shown in the figure, it should be understood that, Figure 2 The examples in the figure are only used to understand the present solution, and the number of terminal devices and servers should be determined flexibly according to actual conditions.

[0082] ​Since the method provided by the embodiments of the present application can be implemented based on the field of artificial intelligence, before introducing the method for recommending information provided by the embodiments of the present application, some basic concepts in the field of artificial intelligence are introduced. AI is a theory, method, technology and application system for simulating, extending and expanding human intelligence by using a digital computer or a machine controlled by a digital computer, perceiving an environment, acquiring knowledge and using the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that machines have the functions of perception, reasoning and decision-making. Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, including both hardware and software technologies. Artificial intelligence basic technologies generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics, etc. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, etc.

[0083] With the research and progress of artificial intelligence technology, artificial intelligence technology is researched in many directions. Machine learning (ML) is a multi-disciplinary subject that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It is a specialized study of how computers simulate or implement human learning behavior to acquire new knowledge or skills, reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rule-based learning.

[0084] Machine learning (ML) is a multi-disciplinary subject that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It is a specialized study of how computers simulate or implement human learning behavior to acquire new knowledge or skills, reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rule-based learning.

[0085] Therefore, the training processes of multiple network models that can be used by the embodiments of the present application will be introduced below.

[0086] (I) First model:

[0087] The first model is used to analyze relevant features of a user, such as user portrait features, exposure click features, and the like, to estimate the exposure rate of a certain push information to the user. Specifically, the first model can be a binary classifier, that is, the output result of the first model can be classified into exposed push information and not exposed push information. When the user portrait features and the exposure click features of a certain user are input into the first model, the first model calculates the probability of exposure of the push information to the user.

[0088] For example, the user portrait features can include the age, gender, occupation, and the like of a certain registered user of a social platform, which will affect the exposure rate of the advertisement. For example, the exposure rate of an advertisement for promoting women's cosmetics in a female user group will be much higher than that in a male user group, and the exposure rate of an advertisement for promoting luxury goods in a user group of 25 years old and above will be much higher than that in a user group of minors. Therefore, by analyzing the user portrait features corresponding to the user, the exposure rate of the advertisement to the candidate user can be predicted, which provides a condition for subsequent screening of the candidate user.

[0089] For example, the exposure click features can include the historical browsing behavior of a certain registered user of a social platform. For example, the login duration of a certain registered user on the social platform in the past week, or the browsing behavior of the registered user on other advertisements similar to the type of the to-be-pushed advertisement in the past year, or the browsing behavior of the registered user on multiple advertisements on a certain advertisement position (the advertisement position corresponding to the to-be-pushed advertisement) in the past month, and the like, which will affect the exposure rate of the advertisement. For example, if a certain registered user has not logged in to the social platform for a long time, the exposure rate of the to-be-pushed advertisement to the user will be greatly reduced. For example, multiple advertisements of the same type as the to-be-pushed advertisement have already been exposed to a certain user on the terminal device corresponding to the user, so the exposure rate of the to-be-pushed advertisement to the user will be greatly improved. For example, a certain registered user is used to browsing and clicking advertisements in a pop-up box, so the exposure rate of the advertisement displayed in the pop-up box to the user will be greatly improved. By analyzing the exposure click features corresponding to the user, the exposure rate of the advertisement to the candidate user can be predicted, which provides a condition for subsequent screening of the candidate user.

[0090] How to train the first model for outputting the exposure rate will be introduced below. Please refer to Figure 3 , Figure 3 A flowchart of the first model training method in the embodiments of the present application is shown in the figure. The first model training flow based on artificial intelligence provided by the present application can include the following steps:

[0091] 301, determine historical recommendation information of the same type as the to-be-recommended information.

[0092] It can be understood that the first model is used to predict the exposure rate of the to-be-recommended information to a certain candidate group. To obtain the first model, the first model can be trained by the data of the historical recommended information of the same type as the to-be-recommended information, that is, the first model needs to learn relevant experience from the recommendation process of the historical recommended information, so as to more accurately predict the exposure rate of the to-be-recommended information. Specifically, the industry information, the delivery position information and the image-text content information corresponding to the to-be-recommended information can be used to determine the similar historical recommended information.

[0093] For example, when an advertiser requests a social platform to push a target advertisement, the advertiser needs to inform the social platform of the industry to which the advertiser belongs, the delivery position to which the advertisement needs to be delivered, and the specific promotion content of the advertisement. Among them, the industry to which the advertiser belongs determines the type of advertisement promotion content. For example, if the industry to which an advertiser belongs is the video application industry, the advertisement pushed by the advertiser will most likely be a video push. If the industry to which an advertiser belongs is the medical equipment industry, the advertisement pushed by the advertiser will most likely be a medical equipment product. Therefore, when the similar historical recommended advertisement to the to-be-recommended advertisement is determined, the historical recommended advertisement can be determined in the historical advertisements of the same industry as the to-be-recommended advertisement according to the industry corresponding to the to-be-recommended advertisement. For example, the social platform can provide an advertisement information table, which can divide the industry into first-level industries and second-level industries. Each first-level industry includes a plurality of second-level industries. When the historical recommended information of the same type as the to-be-recommended information is determined, it can be ensured that the to-be-recommended information and the historical recommended information belong to the same second-level industry.

[0094] After screening a large number of historical advertisements through the industry information corresponding to the to-be-recommended information, the advertisements can be further screened through the delivery position information corresponding to the to-be-recommended information. It can be understood that the social platform can provide a plurality of information delivery positions for information push. Taking advertisement push as an example, the social platform can deliver advertisements on the top position of the social platform page, as shown in Figure 4A , can deliver advertisements through pop-up windows, as shown in Figure 4B , and can deliver advertisements at the opening screen of the social platform, as shown in Figure 4C . It can be understood that different advertisement positions have different delivery effects, and the advertiser can select a suitable advertisement position for advertisement delivery according to the needs of the advertiser. Therefore, to predict the delivery effect of the to-be-recommended advertisement, the historical recommended advertisement needs to be selected which has the same delivery position as the to-be-recommended advertisement. For example, when the to-be-recommended advertisement is delivered in the form of a pop-up window, the historical recommended advertisement similar to the to-be-recommended advertisement needs to be selected from the advertisements displayed in the pop-up window.

[0095] When the server screens out the candidate advertisements corresponding to the same industry information and delivery position information as the to-be-recommended advertisement, the final historical recommended advertisement is also selected from the candidate advertisements according to the specific content of the to-be-recommended advertisement. For example, the text and image information included in the to-be-recommended advertisement can be feature extracted to obtain a corresponding first feature vector, then the text and image information of the candidate advertisement is feature extracted to obtain a corresponding second feature vector, then the similarity between the first feature vector and the second feature vector is calculated through the cosine similarity algorithm, and the final one or more historical recommended advertisements are determined according to the order from high to low of the similarity.

[0096] 302, obtain the people-oriented data and exposure people data corresponding to the historical recommended information.

[0097] When the historical recommended information of the same type as the to-be-recommended information is determined, the related historical recommended data of the historical recommended information is obtained to train the first model. Among them, the advertiser corresponding to the historical recommended information can be obtained first, and then the historical delivery data of the advertiser is searched in the data engine according to the identifier of the advertiser to determine the people-oriented data and exposure people data of the historical recommended information. Among them, the people-oriented data of the historical recommended information refers to the information of all users to which the historical recommended information is pushed, and the exposure people data refers to the user set exposed to the historical recommended information among the users included in the people-oriented data.

[0098] When the people-oriented data and exposure people data corresponding to the historical recommended information are determined, the unexposed people data in the people-oriented data can be determined through the difference set of the two.

[0099] 303, determine the first positive sample training data and the first negative sample training data according to the people-oriented data and the exposure people data corresponding to the historical recommended information.

[0100] Among them, the first positive sample training data can be the exposure people data, wherein the exposure people data includes a plurality of historical users, each historical user carries the user portrait feature and exposure click feature of the user, and the label corresponding to the user, and the label is used to mark the user as an exposed user. In this way, each user in the first positive sample training can be determined as a training data, which is input into the first model in turn to train the first model.

[0101] Among them, the first negative sample training data can be the unexposed people data, and similarly, the unexposed people data also includes a plurality of historical users, each historical user carries the user portrait feature and exposure click feature of the user, and the label corresponding to the user, and the label is used to mark the user as an unexposed user. In this way, each user in the first negative sample training can be determined as a training data, which is input into the first model in turn to train the first model.

[0102] For example, the ratio of the first positive sample training data and the first negative sample training data can be 1 to 2, and the number of positive sample training data is not more than 1 million in order to speed up the training process. At the same time, in order to enhance the robustness of the first model, 10% of the users in the negative sample training data can be randomly sampled from other users of the social platform to join the first negative sample training data, which can improve the training effect of the first model and obtain a more confident first model.

[0103] 304、According to the first positive sample training data and the first negative sample training data, the first model is trained.

[0104] Specifically, the first positive sample training data and the first negative sample training data can be input into the first model in sequence, and the first model can output the classification result of the user (whether the user belongs to the exposure population or the non-exposure population) by analyzing and operating the user portrait features and the exposure click features. Then, the actual classification result of the user is obtained according to the label of the user, the similarity between the actual classification result and the output classification result of the user is evaluated by a loss function, and then the related operation parameters of the first model are updated in reverse iteration according to the loss value. The training process of the first model ends until the training condition is reached.

[0105] For example, the training end condition can be that the corresponding loss value converges, that is, when the loss value converges, the training process of the first model ends. The training end condition can also be that the preset training times are reached, that is, when the training times of the first model reach 1 million, the training process of the first model ends, and the specific form is not limited.

[0106] For example, the cross-entropy loss function can be used as the loss function of the first model. Since the first model can be a binary classifier, the loss function can be:

[0107]

[0108] Where C is the loss value, n represents the total number of samples, x represents the sample, y is the actual label of the xth sample, and a is the data result of the xth sample predicted by the classifier.

[0109] Through the above training process, the first model can be trained by the related data corresponding to the historical recommendation information, so that the output result of the first model can more effectively predict the exposure rate of the to-be-recommended information to the candidate user.

[0110] (II) Second model:

[0111] The second model is used to analyze relevant features of the user, such as user portrait features, exposure click features, etc., to estimate the click rate of a certain push information for the user. Specifically, the second model can also be a binary classifier, that is, the output result of the second model can be divided into clicking the push information and exposing but not clicking the push information. When the user portrait features and exposure click features of a certain user are input into the second model, the second model calculates the probability that the push information is clicked by the user.

[0112] For example, the user portrait features can be the age, gender, occupation, etc. of a registered user of a social platform, which will affect the click rate of the advertisement. For example, female consumer groups are more willing to read multiple advertisements to learn about multiple products, so the click rate of the advertisement will be affected by the gender of the user. For example, adult groups over the age of 20 are more willing to obtain electronic product consultations through the Internet, so the click rate of the electronic product advertisement will be affected by the age of the user. Therefore, by analyzing the user portrait features corresponding to the user, the click rate of the advertisement for the candidate user can be predicted, providing conditions for subsequent screening of the candidate user.

[0113] For example, the exposure click features can be the historical reading behavior of a registered user of a social platform. For example, the length of time a registered user logs into a social platform in the past week, or the reading behavior of the registered user for other advertisements similar to the type of the advertisement to be pushed in the past year, or the reading behavior of the registered user for multiple advertisements on a certain advertisement position (the advertisement position corresponding to the advertisement to be pushed) in the past month, etc., which will affect the click rate of the advertisement. By analyzing the exposure click features corresponding to the user, the click rate of the advertisement for the candidate user can be predicted, providing conditions for subsequent screening of the candidate user.

[0114] The following will introduce how to train the second model for outputting the click rate. Please refer to Figure 5 , Figure 5 A flowchart of the second model training method in the embodiments of the present application is shown in the figure. The second model training flow provided by the present application based on artificial intelligence can include:

[0115] 501, determine historical recommendation information of the same type as the to-be-recommended information.

[0116] For example, this step is similar to step 301 in the embodiment shown in Figure 3 , which will not be repeated here.

[0117] 502, obtain exposure population data and click population data corresponding to the historical recommendation information.

[0118] When the historical recommendation information of the same type as the to-be-recommended information is determined, the related historical recommendation data of the historical recommendation information needs to be obtained to train the second model. Among them, the advertiser corresponding to the historical recommendation information can be obtained first, and then the historical delivery data of the advertiser is searched in the data engine according to the identifier of the advertiser to determine the exposure population data and the click population data in the population targeting data of the historical recommendation information. Among them, the population targeting data of the historical recommendation information refers to the information of all users to which the historical recommendation information is pushed, the exposure population data refers to the set of users in the population targeting data who are exposed to the historical recommendation information, and the click population data refers to the set of users in the exposure population data who have clicked and viewed the historical recommendation information.

[0119] When the exposure population data and the click population data corresponding to the historical recommendation information are determined, the difference set of the two can be used to determine the exposure non-click population data in the exposure population data.

[0120] 503、According to the exposure population data and the click population data corresponding to the historical recommendation information, the second positive sample training data and the second negative sample training data are determined.

[0121] Among them, the first positive sample training data can be the click population data, wherein the click population data includes a plurality of historical users, each historical user carries the user portrait feature and the exposure click feature of the user, and the label corresponding to the user, and the label is used to mark the user as a click user. In this way, each user in the second positive sample training can be determined as a training data, which is input into the second model in turn to train the third model.

[0122] Among them, the first negative sample training data can be the exposure non-click population data, and similarly, the exposure non-click population data also includes a plurality of historical users, each historical user carries the user portrait feature and the exposure click feature of the user, and the label corresponding to the user, and the label is used to mark the user as an exposure non-click user. In this way, each user in the second negative sample training can be determined as a training data, which is input into the second model in turn to train the second model.

[0123] For example, the ratio of the second positive sample training data and the second negative sample training data can be 1:5. In order to speed up the training process, the number of positive sample training data is controlled to be about 500,000.

[0124] 504、According to the second positive sample training data and the second negative sample training data, the second model is trained.

[0125] Specifically, the second positive sample training data and the second negative sample training data can be sequentially input into the second model, the second model analyzes and calculates the user portrait feature and the exposure click feature, and outputs the classification result corresponding to the user (whether the user belongs to the exposure click crowd or the exposure non-click crowd), then the actual classification result is obtained according to the label of the user, the similarity between the actual classification result and the output classification result is evaluated through the loss function, and then the related operation parameters of the second model are updated in reverse iteration according to the loss value, until the training condition is reached, and the training process of the second model ends.

[0126] For example, the training end condition can be that the corresponding loss value converges, that is, when the loss value converges, the training process of the second model ends. The training end condition can also be that the preset training times are reached, that is, when the training times of the second model reach 500,000 times, the training process of the second model ends, and the specific form is not limited.

[0127] For example, the cross-entropy loss function can be used as the loss function of the second model. Since the second model can also be a binary classifier, the loss function can refer to the loss function of the first model shown in the embodiment, which will not be repeated here. Figure 3

[0128] Through the above training process, the second model can be trained by the related data corresponding to the historical recommendation information, so that the output result of the second model can more effectively predict the click rate of the to-be-recommended information for the candidate user.

[0129] (Three) third model:

[0130] The third model is used to analyze the related features of the user, such as the user portrait feature and the exposure click feature, to estimate the click rate of a certain push information for the user. Specifically, the third model can also be a binary classifier, that is, the output result of the third model can be divided into push information conversion and push information click non-conversion. When the user portrait feature of a certain user is input into the third model, the third model calculates the conversion rate of the to-be-recommended information for the user.

[0131] For example, the user portrait feature can be the age, gender, occupation, etc. of a registered user on a social platform, which will affect the conversion rate of the advertisement. For example, users aged 20-30 are more likely to receive online shopping than users over 40, and the probability of purchasing a certain product through an advertisement will be greatly improved. For example, people in the Internet profession are more willing to pay for a paid video, so the occupation of the user can affect the conversion rate of the paid video. Therefore, by analyzing the user portrait feature corresponding to the user, the conversion rate of the advertisement for the candidate user can be predicted, which provides a condition for subsequent screening of the candidate user. ​

[0132] The third model for outputting conversion rate will be introduced as follows. Please refer to Figure 6 , Figure 6 A flowchart of the third model training method in the embodiments of the present application is shown in the figure. The third model training flow based on artificial intelligence provided by the present application can include the following steps:

[0133] 601. Determine historical recommended information of the same type as the to-be-recommended information.

[0134] The exemplary step is similar to step 301 in the embodiment shown in Figure 3 , which will not be repeated here.

[0135] 602. Obtain the click user group data and conversion user group data corresponding to the historical recommended information.

[0136] After determining the historical recommended information of the same type as the to-be-recommended information, the relevant historical recommended data of the historical recommended information needs to be obtained to train the second model. Among them, the ad master corresponding to the historical recommended information can be obtained first, and then the historical delivery data of the ad master is searched in the data engine according to the identifier of the ad master to determine the click user group data and conversion user group data in the user group targeting data of the historical recommended information. Among them, the user group targeting data of the historical recommended information refers to the information of all users pushed by the historical recommended information. The click user group data is the set of users in the user group targeting data who have clicked and viewed the historical recommended information. The conversion user group data refers to the set of users in the click user group data who have purchased related products through the link in the recommended information.

[0137] After determining the click user group data and conversion user group data corresponding to the historical recommended information, the difference set of the two can be used to determine the click non-conversion user group data in the click user group data.

[0138] 603. Determine the third positive sample training data and the third negative sample training data according to the click user group data and the conversion user group data corresponding to the historical recommended information.

[0139] Among them, the third positive sample training data can be the conversion user group data, wherein the conversion user group data includes a plurality of historical users, each historical user carries the user portrait feature of the user and the label corresponding to the user, and the label is used to mark the user as a conversion user. In this way, each user in the third positive sample training can be determined as a training data, which is input into the third model in turn to train the third model.

[0140] The third negative sample training data can be click non-conversion population data. Similarly, the click non-conversion population data also includes a plurality of historical users, each of which carries a user portrait feature of the user and a label corresponding to the user, and the label is used to mark the user as a click non-conversion user. In this way, each user in the third negative sample training can be determined as a training data and sequentially input into the third model to train the third model.

[0141] 604. Training the third model according to the third positive sample training data and the third negative sample training data.

[0142] Specifically, the third positive sample training data and the third negative sample training data can be sequentially input into the third model. The third model analyzes and operates the user portrait feature, outputs a classification result (whether the user belongs to the conversion population or the click non-conversion population) corresponding to the user, then acquires an actual classification result according to the label of the user, evaluates the similarity of the actual classification result and the output classification result through a loss function, and then iteratively updates the related operation parameters of the third model according to the loss value until the training condition is reached, and the training process of the third model ends.

[0143] Exemplarily, since the third model can also be a binary classifier, the loss function and the training condition corresponding to the third model can refer to the loss function and the training condition of the first model shown in the embodiment. Figure 3 The first model in the embodiment will not be described herein.

[0144] Through the above training process, the third model can be trained by the related data corresponding to the historical recommendation information, so that the output result of the third model can more effectively predict the conversion rate of the to-be-recommended information for the candidate user.

[0145] (Four) Information recommendation estimation model:

[0146] The information recommendation estimation model can be a multi-classification model, which fuses and scores the exposure rate, click rate and conversion rate input into the model. First, the estimated exposure rate, estimated click rate and estimated conversion rate of the to-be-recommended information for the user are estimated according to the user portrait feature of the user, and then the estimated exposure rate, estimated click rate and estimated conversion rate are used as the user feature of the user. Then, the user feature is input into the information recommendation estimation model to obtain the final delivery effect of the recommended information for the user.

[0147] Exemplarily, the information recommendation prediction model can be a multi-classifier, wherein the output result can be divided into four categories, i.e., a converted user, a clicked non-converted user, an exposed non-clicked user, and a non-exposed user, and the output layer can adopt a softmax function. When the server inputs the user features (estimated exposure rate, estimated click rate, and estimated conversion rate) corresponding to a user into the information recommendation prediction model, the information recommendation prediction model fuses and scores the user features, and outputs the probability of the user belonging to each of the above four categories.

[0148] Exemplarily, the information recommendation prediction model can also be a scoring model, i.e., when the server inputs the user features (estimated exposure rate, estimated click rate, and estimated conversion rate) corresponding to a user into the information recommendation prediction model, the information recommendation prediction model analyzes the user features, and outputs a delivery score corresponding to the user. It can be understood that the delivery score is an evaluation of whether the recommended information is suitable for being delivered to the user. Specifically, the network structure of the information recommendation prediction model is not limited.

[0149] Next, how to train the information recommendation prediction model for outputting the delivery result of the recommended model will be introduced. Please refer to Figure 7 , Figure 7 A flowchart of the information recommendation prediction model training method in the embodiments of the present application is shown in the figure. The information recommendation prediction model flow based on artificial intelligence provided by the present application can include the following steps.

[0150] 701. Determine historical recommended information of the same type as the recommended information.

[0151] Exemplarily, this step is similar to step 301 in the embodiment shown in Figure 3 , and will not be repeated here.

[0152] 702. Obtain people-oriented data corresponding to the historical recommended information.

[0153] It can be understood that the information recommendation prediction model also needs to predict the data of the historical recommended information of the same type as the recommended information. The people-oriented data corresponding to the historical recommended information includes a plurality of historical users, each historical user corresponds to user portrait features and exposure and click features, and each historical user carries a label, which is used to indicate the actual delivery result of the historical recommended information to a certain historical user. For example, if the label of a certain historical user is exposure and non-click, it means that when the historical recommended information is recommended to the historical user, the information is exposed on the terminal device used by the historical user, but is not clicked. Each user in the people-oriented data can be determined as a sample, and each sample is input into the information recommendation prediction model in turn to train the model.

[0154] 703、According to the user portrait features and exposure and click features corresponding to each user in the crowd orientation data, obtain the estimated exposure rate, estimated click rate and estimated conversion rate of the historical recommendation information for each user.

[0155] Before inputting each sample into the information recommendation estimation model, the estimated exposure rate, estimated click rate and estimated conversion rate of the historical recommendation information for each user need to be obtained according to the user portrait features and exposure and click features of the historical user included in each sample. Exemplarily, the estimated exposure rate can be obtained by the first model, the estimated click rate can be obtained by the second model, and the estimated conversion rate can be obtained by the second model, and then the training sample is determined according to the estimated exposure rate, estimated click rate and estimated conversion rate.

[0156] 704、According to the estimated exposure rate, estimated click rate and estimated conversion rate of the historical recommendation information for each user, determine the training sample data.

[0157] As can be known from the above description, each training sample corresponds to a historical user, and each training sample includes the estimated exposure rate, estimated click rate and estimated conversion rate of the historical user, and the user label. When a plurality of training samples are determined, the information recommendation estimation model can be trained by the training samples. It can be understood that, in order to ensure balanced sampling, part of the training data needs to be down-sampled, that is, the number of training samples carrying the converted user label, the number of training samples carrying the clicked non-converted user label, the number of training samples carrying the exposed non-clicked user label and the number of training samples carrying the non-exposed user label are relatively balanced.

[0158] 705、According to the training sample data, train the information recommendation estimation model.

[0159] Specifically, the features of the determined training samples can be input into the information recommendation estimation model in sequence. The information recommendation estimation model analyzes and calculates the estimated exposure rate, estimated click rate and estimated conversion rate, outputs the classification result corresponding to the user, then obtains the actual classification result according to the label of the user, evaluates the similarity between the actual classification result and the output classification result through the loss function, then updates the related operation parameters of the recommendation estimation model according to the loss value, and the training process of the recommendation estimation model ends until the training condition is reached.

[0160] Exemplarily, the training end condition can be that the corresponding loss value converges, that is, when the loss value converges, the training process of the information recommendation estimation model ends. The training end condition can also be that the preset training times are reached, that is, when the training times of the information recommendation estimation model reach 500,000 times, the training process of the information recommendation estimation model ends, and the specific form is not limited.

[0161] Exemplarily, a cross-entropy loss function can be used as the loss function of the information recommendation estimation model. Since the information recommendation estimation model can be a four-classifier, the loss function can be:

[0162]

[0163] wherein J is the loss value, K is the number of label categories, i.e., if it is a four-classifier, K is 4. y is the label, p i is the probability that the category is i.

[0164] Through the above training process, the information recommendation estimation model can be trained by the relevant data corresponding to the historical recommendation information, so that the output result of the information recommendation estimation model can more effectively predict the actual delivery result of the to-be-recommended information to the candidate user.

[0165] In combination with the above description, before introducing the scheme provided by the embodiments of the present application, the system framework corresponding to the information recommendation method is introduced, Figure 8 A structural schematic diagram of an information recommendation system provided by the embodiments of the present application includes a first model, a second model, a third model and an information recommendation estimation model. The first model, the second model and the third model are connected with the information recommendation estimation model. When the server determines the targeting crowd for the to-be-recommended information, it is first determined whether the advertiser corresponding to the to-be-recommended information is a new advertiser. If it is a new advertiser, similar advertisers are found, and the candidate crowd is determined from the historical advertisements pushed by the similar advertisers. If it is an old advertiser, the candidate crowd is determined from the historical users pushed by the advertiser. Then the user portrait features and the exposure and click features of each candidate user in the candidate crowd are determined, and then they are input into the first model, the second model and the third model respectively to obtain the estimated exposure rate, the estimated click rate and the estimated conversion rate corresponding to the candidate user respectively. Then the three are input into the information recommendation estimation model to obtain the final delivery result for the candidate user.

[0166] Figure 9 A flowchart of an information recommendation method provided by the embodiments of the present application is shown in the figure. The information recommendation method includes the following steps:

[0167] 901、Receiving an intelligent targeting request for a to-be-recommended information.

[0168] Taking the advertisement delivery as an example, when an advertiser needs to deliver an advertisement to a social platform, the advertiser first needs to send an intelligent targeting request to the server to require the server to recommend appropriate candidate crowds as the targeting crowd of the to-be-recommended information. Specifically, the server can provide an operation page for the advertiser, and the advertiser sends an intelligent targeting request to the server through the operation page, such as Figure 10AAs shown, the operation page can include the account of the advertiser and a related parameter setting interface.

[0169] The advertiser account can be used to obtain information of the advertiser, including industry information of the advertiser, advertisement placement information, and graphic information of the advertisement to be placed by the advertiser. The parameter setting page includes other requirements of the advertiser, such as the update mode when the server provides services for the advertiser, the optimization target, and the like. For example, when the advertiser selects the update mode as "model estimation correction" and the optimization target as "conversion rate" in the update setting, the server needs to output a targeted number package with a high advertisement conversion rate every day.

[0170] For example, the present scheme can also be applied to the automatic expansion module of the placement end of the social platform. When the advertiser enables the automatic expansion function on the page as shown, Figure 10B The social platform will start the intelligent targeting function to continuously generate new targeted groups for the advertiser. It can be understood that this way can improve the starting speed of the automatic expansion module and optimize the advertisement conversion quantity after the advertiser uses the automatic expansion module.

[0171] 902. Determine the candidate user according to the intelligent targeting request.

[0172] When the server receives the intelligent targeting request, it needs to select candidate users from the large pool of users according to the intelligent targeting request to determine whether to push the to-be-recommended information to the candidate users. For example, in the intelligent targeting request, the advertiser requires that the size of the targeted number package is 100,000 users, so the intelligent targeting request can select 200,000 users from the large pool of users to determine the final targeted group.

[0173] For example, according to the industry of the advertiser in the intelligent targeting request, the candidate users can be determined, for example, the industry of a certain advertiser is a cosmetic manufacturer, so the server can select candidate users from female users according to the nature of the industry in order to determine the subsequent targeted group.

[0174] For example, the candidate users can also be selected according to the graphic content of the advertisement in the intelligent targeting request, for example, a plurality of advertisements of the same type are determined according to the graphic content of the advertisement, the candidate users are selected from the placement groups of these advertisements, and then the candidate users are screened to obtain the final targeted group.

[0175] 903. Obtain the user portrait features and exposure click features corresponding to the candidate users.

[0176] When the candidate users are circled, the user portrait features and exposure click features of the candidate users need to be obtained. For example, the user portrait features of the candidate users can include age, gender, occupation, etc. These features will affect the advertising effect. Therefore, by analyzing the user portrait features of the candidate users, the advertising effect of the to-be-recommended information on the candidate users can be predicted, thereby providing a condition for subsequent screening of the candidate users.

[0177] For example, the exposure click features of the candidate users can include historical browsing behaviors. For example, the candidate user logs in the social platform for a time period within the past week, or the candidate user browses other advertisements similar to the type of the to-be-pushed advertisement within the past year, etc. These will also affect the advertising effect. By analyzing the exposure click features of the candidate users, the advertising effect of the to-be-recommended information on the candidate users can be predicted, thereby providing a condition for subsequent screening of the candidate users.

[0178] 904. According to the user portrait features and the exposure click features of the candidate users, an estimated exposure rate, an estimated click rate, and an estimated conversion rate of the to-be-recommended information on the candidate users are determined.

[0179] For example, the user portrait features and the exposure click features of the candidate users can be input into a first model to determine the estimated exposure rate of the to-be-recommended information on the candidate users by the first model. The user portrait features and the exposure click features of the candidate users can be input into a second model to determine the estimated click rate of the to-be-recommended information on the candidate users by the second model. The user portrait features of the candidate users can be input into a third model to determine the estimated conversion rate of the to-be-recommended information on the candidate users by the third model.

[0180] It can be understood that the first model, the second model, and the third model are each a model as described above, and the training process is also as described above. Figures 3 to 6 Therefore, details are not repeated here.

[0181] 905. According to the estimated exposure rate, the estimated click rate, and the estimated conversion rate, an estimated advertising result of the to-be-recommended information on the candidate users is determined.

[0182] For example, the estimated exposure rate, the estimated click rate, and the estimated conversion rate of the candidate users can be input into an information recommendation estimation model to output the estimated advertising result of the to-be-recommended information on the candidate users by the information recommendation estimation model. It can be understood that the information recommendation estimation model is the information recommendation estimation model in the embodiment as described above, and details are not repeated here. Figure 7

[0183] 906. According to the estimated advertising result, the to-be-recommended information is recommended to a terminal device used by the candidate user.

[0184] ​When the information recommendation estimation model outputs the predicted delivery result corresponding to each candidate user, the targeted population can be selected from the candidate users according to the predicted delivery result. For example, when the information recommendation estimation model outputs the classification of the candidate users, the targeted population can be selected from the candidate users whose output result is the conversion population. If the information recommendation estimation model outputs the delivery score corresponding to the candidate user, the candidate users with high delivery scores need to be selected for advertisement delivery according to the size of the delivery score.

[0185] In the embodiments of the present application, a method for information recommendation is provided. First, the exposure rate, the click rate and the conversion rate of the to-be-recommended information with respect to the candidate user are estimated according to the user portrait features and the exposure-click features of the candidate user. Then, the final delivery result of the to-be-recommended information with respect to the user is predicted by analyzing the estimated exposure rate, the estimated click rate and the estimated conversion rate. In this way, the user with better delivery result can be selected for information delivery according to the predicted delivery result, so as to achieve the effect of accurate delivery of the to-be-recommended information. In this way, the useless information delivery behavior can be reduced, thereby reducing a large amount of redundant information in the Internet, saving network resources, improving the conversion rate of the to-be-recommended information and reducing the information delivery cost.

[0186] Figure 11 A structural schematic diagram of an information recommendation device is provided for the embodiments of the present application, which includes:

[0187] The acquisition unit 1101 is configured to acquire the user portrait features corresponding to the candidate user of the to-be-recommended information.

[0188] The determination unit 1102 is configured to determine the exposure-click features corresponding to the candidate user according to the information recommendation state of the recommended information of the candidate user in the historical time period, wherein the recommended information in the historical time period has an association relationship with the to-be-recommended information.

[0189] The determination unit 1102 is further configured to determine the estimated exposure rate, the estimated click rate and the estimated conversion rate of the to-be-recommended information with respect to the candidate user according to the user portrait features and the exposure-click features of the candidate user.

[0190] The processing unit 1103 is configured to determine the estimated delivery result of the to-be-recommended information with respect to the candidate user according to the estimated exposure rate, the estimated click rate and the estimated conversion rate of the to-be-recommended information with respect to the candidate user.

[0191] The sending unit 1104 is configured to recommend the to-be-recommended information to the terminal device used by the candidate user if the estimated delivery result meets the information delivery condition.

[0192] In a possible design, the determining unit 1102 is specifically configured to input the user portrait feature and the exposure click feature of the candidate user into the first model, and determine, by the first model, the estimated exposure rate of the to-be-recommended information for the candidate user. The user portrait feature and the exposure click feature of the candidate user are input into the second model, and the click rate of the to-be-recommended information for the candidate user is determined by the second model. The user portrait feature of the candidate user is input into the third model, and the estimated conversion rate of the to-be-recommended information for the candidate user is determined by the third model.

[0193] In a possible design, the information recommendation apparatus further includes a training unit 1105, and the training unit 1105 is configured to:

[0194] The obtaining unit 1101 is further configured to obtain population orientation data and exposure population data corresponding to at least one historical recommended information, where the information type of the historical recommended information is the same as that of the to-be-recommended information.

[0195] The determining unit 1102 is further configured to determine, according to the population orientation data and the exposure population data, unexposed population data corresponding to the historical recommended information. The exposure population data is determined as first positive sample training data, and the unexposed population data is determined as first negative sample training data, where each first positive sample training data and each first negative sample training data include a user portrait feature and an exposure click feature of a historical user, and an actual exposure result of the historical recommended information for the historical user.

[0196] The training unit 1105 is configured to train the first model according to the first positive sample training data and the first negative sample training data.

[0197] In a possible design, the information recommendation apparatus further includes a training unit 1105, and the training unit 1105 is configured to:

[0198] The obtaining unit 1101 is further configured to obtain exposure population data and click population data corresponding to at least one historical recommended information, where the information type of the historical recommended information is the same as that of the to-be-recommended information.

[0199] The determining unit 1102 is further configured to determine, according to the exposure population data and the click population data, exposure non-click population data corresponding to the historical recommended information. The click population data is determined as second positive sample training data, and the exposure non-click population data is determined as second negative sample training data, where each second positive sample training data and each second negative sample training data include a user portrait feature and an exposure click feature of a historical user, and an actual click result of the historical recommended information for the historical user.

[0200] The training unit 1105 is configured to train the second model according to the second positive sample training data and the second negative sample training data.

[0201] In a possible design, the information recommendation apparatus further includes a training unit 1105, and wherein:

[0202] The obtaining unit 1101 is configured to obtain click user group data and conversion user group data corresponding to at least one historical recommendation information, wherein the information type of the historical recommendation information and the to-be-recommended information is same.

[0203] The determining unit 1102 is configured to determine click non-conversion user group data corresponding to the historical recommendation information according to the click user group data and the conversion user group data, determine the conversion user group data as third positive sample training data, and determine the click non-conversion user group data as third negative sample training data, wherein each third positive sample training data and each third negative sample training data include user portrait features of a historical user and an actual conversion result of the historical recommendation information to the historical user.

[0204] The training unit 1105 is configured to train a third model according to the third positive sample training data and the third negative sample training data.

[0205] In a possible design, the determining unit 1102 is specifically configured to input an estimated exposure rate, an estimated click rate and an estimated conversion rate of the to-be-recommended information to the candidate user into the information recommendation estimation model, and output an estimated delivery result of the to-be-recommended information to the candidate user by the information recommendation estimation model.

[0206] In a possible design, the information recommendation apparatus further includes a training unit 1105, and wherein:

[0207] The obtaining unit 1101 is further configured to obtain user group targeting data corresponding to at least one historical recommendation information, the information type of the historical recommendation information and the to-be-recommended information is same, the user group targeting data includes user portrait features and exposure click features of a historical user, and an actual delivery result of the historical recommendation information to the historical user.

[0208] The obtaining unit 1101 is further configured to input the user portrait features and the exposure click features into a first model, and obtain an estimated exposure rate of the historical recommendation information to the historical user by the first model. The obtaining unit 1101 is further configured to input the user portrait features and the exposure click features into a second model, and obtain an estimated click rate of the historical recommendation information to the historical user by the second model. The obtaining unit 1101 is further configured to input the user portrait features into a third model, and obtain an estimated conversion rate of the historical recommendation information to the historical user by the third model.

[0209] The determining unit 1102 is further configured to input the estimated exposure rate, the estimated click rate and the estimated conversion rate of the historical recommendation information to the historical user into an information recommendation estimation model, and obtain an estimated delivery result of the historical recommendation information to the historical user by the information recommendation estimation model.

[0210] The training unit 1105 is configured to train the information recommendation estimation model according to a loss function, where the loss function is used to evaluate the similarity between the actual delivery result and the predicted delivery result of the historical recommendation information for the historical user.

[0211] In a possible design, the determining unit 1102 is specifically configured to classify the candidate user by using the information recommendation estimation model.

[0212] The sending unit 1104 is specifically configured to recommend the to-be-recommended information to a terminal device used by the candidate user if the candidate user belongs to the target classified population.

[0213] In a possible design, the determining unit 1102 is specifically configured to determine the delivery score of the to-be-recommended information for the candidate user by using the information recommendation estimation model.

[0214] The sending unit 1104 is specifically configured to recommend the to-be-recommended information to a terminal device used by the candidate user if the delivery score exceeds a preset threshold.

[0215] In a possible design, the obtaining unit 1101 is further configured to obtain a first information recommendation state of the candidate user for a target information push position in a historical time period, where the first information recommendation state includes a push state of at least one recommended information on the target information push position, and the target information push position is an information push position corresponding to the to-be-recommended information.

[0216] The determining unit 1102 is specifically configured to determine, according to the first information recommendation state, an information exposure number and an information click number of the at least one recommended information on the target information push position, and determine the information exposure number and the information click number of the candidate user on the target information push position as exposure and click features corresponding to the candidate user.

[0217] In a possible design, the obtaining unit 1101 is further configured to obtain a second information recommendation state of the candidate user for a target recommended information in a historical time period, where the second information recommendation state includes a push state of the target recommended information on at least one information push position, and the target recommended information is of the same type as the to-be-recommended information.

[0218] The determining unit 1102 is specifically configured to determine, according to the second information recommendation state, an information exposure number and an information click number of the target recommended information on the at least one information push position, and determine the information exposure number and the information click number of the target recommended information on the at least one information push position as exposure and click features corresponding to the candidate user.

[0219] In a possible design, the obtaining unit 1101 is further configured to obtain state information corresponding to the to-be-recommended information, where the state information includes industry information, target delivery position information and image-text content information corresponding to the to-be-recommended information.

[0220] The acquisition unit 1101 is further used to acquire candidate information according to the status information corresponding to the information to be recommended, wherein the industry information corresponding to the candidate information is the same as the industry information corresponding to the information to be recommended, and the target placement information corresponding to the candidate information is the same as the target placement information corresponding to the information to be recommended.

[0221] The determination unit 1102 is also used to obtain a first feature vector corresponding to the information to be recommended based on the graphic content information corresponding to the information to be recommended, obtain a second feature vector corresponding to the candidate information based on the graphic content information corresponding to the candidate information, and determine the historical recommendation information corresponding to the information to be recommended from the candidate information based on the cosine similarity between the first feature vector and the second feature vector.

[0222] The present application also provides another information recommendation device that can be deployed on a server. Figure 12 , Figure 12 This is a schematic diagram of an embodiment of a server in an embodiment of the present application. As shown in the figure, the server 1200 may have relatively large differences due to different configurations or performances, and may include one or more central processing units (CPUs) 1222 (for example, one or more processors) and memory 1232, and one or more storage media 1230 (for example, one or more mass storage devices) for storing application programs 1242 or data 1244. Among them, the memory 1232 and the storage medium 1230 can be temporary storage or permanent storage. The program stored in the storage medium 1230 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations on the server. Furthermore, the central processing unit 1222 can be configured to communicate with the storage medium 1230 to execute a series of instruction operations in the storage medium 1230 on the server 1200.

[0223] The server 1200 may also include one or more power supplies 1226, one or more wired or wireless network interfaces 1250, one or more input and output interfaces 1258, and / or one or more operating systems 1241, such as Windows Server 200. TM , Mac OS X TM , Unix TM , Linux TM , FreeBSD TM etc.

[0224] The steps performed by the server in the above embodiment can be based on the Figure 12 The server structure shown.

[0225] In the embodiment of the present application, the CPU 1222 included in the server is used for the following steps:

[0226] Obtain user profile features corresponding to candidate users for whom information is to be recommended.

[0227] The exposure click feature corresponding to the candidate user is determined according to the information recommendation status of the recommended information of the candidate user in the historical time period, wherein the recommended information in the historical time period has an associated relationship with the information to be recommended.

[0228] Based on the user portrait characteristics and exposure-click characteristics of the candidate users, the estimated exposure rate, estimated click rate and estimated conversion rate of the recommended information for the candidate users are determined.

[0229] The estimated delivery result of the information to be recommended for the candidate user is determined based on the estimated exposure rate, estimated click rate and estimated conversion rate of the information to be recommended for the candidate user.

[0230] If the estimated delivery result meets the information delivery conditions, the information to be recommended is recommended to the terminal device used by the candidate user.

[0231] The present application also provides a computer-readable storage medium in which a computer program is stored. When the computer program is run on a computer, the computer executes the above-mentioned Figures 3 to 9 The illustrated embodiment describes the steps performed by the server in the method.

[0232] The present application also provides a computer program product or computer program including a program, wherein the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the aforementioned Figures 3 to 9 The illustrated embodiment describes the steps performed by the server in the method.

[0233] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0234] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the described device embodiments are merely schematic. For example, the division of the units is only a logical function division. There can be another division manner for the actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different units, can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0235] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0236] In addition, each functional unit in the various embodiments of the present application can be integrated into a processing unit, or each unit can exist physically, or two or more units can be integrated into one unit. The integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0237] The integrated unit, if implemented in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0238] The above-described embodiments are merely used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for information recommendation, characterized in that: The method comprises: Obtain user profile features corresponding to candidate users for whom information is to be recommended; Determining the exposure-click feature corresponding to the candidate user based on the information recommendation status of the recommended information of the candidate user in a historical time period, where the recommended information in the historical time period is of the same or similar type as the information to be recommended; the information recommendation status is the push status of the recommended information in the information push position; the exposure-click feature includes the number of information exposures and the number of information clicks; Determine, based on the user portrait features and exposure-click features of the candidate user, an estimated exposure rate, an estimated click rate, and an estimated conversion rate of the information to be recommended for the candidate user; Determining an estimated delivery result of the information to be recommended for the candidate user based on the estimated exposure rate, the estimated click-through rate, and the estimated conversion rate of the information to be recommended for the candidate user; If the estimated delivery result meets the information delivery condition, the information to be recommended is recommended to the candidate user.

2. The method according to claim 1, characterized in that Determining the estimated exposure rate, estimated click rate, and estimated conversion rate of the information to be recommended for the candidate user based on the user portrait feature and the exposure-click feature of the candidate user includes: Inputting the user portrait features and the exposure-click features of the candidate user into a first model, and determining an estimated exposure rate of the to-be-recommended information for the candidate user through the first model; Inputting the user portrait features and the exposure-click features of the candidate user into a second model, and determining an estimated click-through rate of the information to be recommended for the candidate user through the second model; The user portrait features of the candidate user are input into a third model, and the estimated conversion rate of the information to be recommended for the candidate user is determined by the third model.

3. The method according to claim 2, characterized in that The method further comprises: Obtaining demographic data and exposure demographic data corresponding to at least one historical recommendation information; the historical recommendation information and the information to be recommended are of the same information type; Determining the unexposed population data corresponding to the historical recommendation information according to the population targeting data and the exposed population data; Determine the exposed population data as first positive sample training data, and determine the unexposed population data as first negative sample training data; wherein each first positive sample training data and each first negative sample training data includes user portrait features and exposure click features of historical users, as well as actual exposure results of the historical recommendation information for the historical users; The first model is trained based on the first positive sample training data and the first negative sample training data.

4. The method according to claim 2, characterized in that The method further comprises: Obtaining exposure population data and click population data corresponding to at least one historical recommendation information; the historical recommendation information and the information to be recommended are of the same information type; Determine the exposed but not clicked population data corresponding to the historical recommendation information according to the exposed population data and the clicked population data; Determine the click population data as second positive sample training data, and determine the exposure-but-non-click population data as second negative sample training data; wherein each second positive sample training data and each second negative sample training data includes user portrait features and exposure-click features of historical users, as well as actual click results of the historical recommendation information for the historical users; The second model is trained based on the second positive sample training data and the second negative sample training data.

5. The method according to claim 2, characterized in that The method further comprises: Obtaining click population data and conversion population data corresponding to at least one historical recommendation information; the historical recommendation information and the information to be recommended are of the same information type; Determine the click-unconverted population data corresponding to the historical recommendation information according to the click population data and the converted population data; Determine the conversion population data as third positive sample training data, and determine the click-unconverted population data as third negative sample training data; wherein each third positive sample training data and each third negative sample training data includes user portrait features of historical users, and actual conversion results of the historical recommendation information for the historical users; The third model is trained according to the third positive sample training data and the third negative sample training data.

6. The method according to any one of claims 2 to 5, characterized in that Determining an estimated delivery result of the information to be recommended for the candidate user based on the estimated exposure rate, the estimated click rate, and the estimated conversion rate of the information to be recommended for the candidate user includes: Outputting an estimated delivery result of the information to be recommended for the candidate user through an information recommendation estimation model; the information recommendation estimation model is used to obtain a classification result or delivery score based on the estimated exposure rate, the estimated click rate, and the estimated conversion rate of the candidate user; The input end of the information recommendation estimation model is connected to the output end of the first model, the output end of the second model and the output end of the third model.

7. The method according to claim 6, characterized in that The method further comprises: Obtaining demographic data corresponding to at least one historical recommendation information, where the historical recommendation information and the information to be recommended are of the same information type; the demographic data includes user portrait features and exposure and click features of the historical user, as well as actual delivery results of the historical recommendation information for the historical user; Inputting the user portrait feature and the exposure click feature into a first model, and obtaining an estimated exposure rate of the historical recommendation information for the historical user through the first model; Inputting the user portrait feature and the exposure-click feature into a second model, and obtaining an estimated click rate of the historical recommendation information for the historical user through the second model; Inputting the user portrait features into a third model, and obtaining an estimated conversion rate of the historical recommendation information for the historical user through the third model; Inputting the estimated exposure rate, estimated click rate, and estimated conversion rate of the historical recommendation information for the historical user into the information recommendation estimation model, and obtaining the estimated delivery result of the historical recommendation information for the historical user through the information recommendation estimation model; The information recommendation prediction model is trained according to a loss function, wherein the loss function is used to evaluate the similarity between the actual delivery result and the predicted delivery result of the historical recommendation information for the historical user.

8. The method according to claim 6, characterized in that Outputting the estimated delivery result of the information to be recommended for the candidate user through the information recommendation estimation model includes: Classifying the candidate users by using the information recommendation prediction model; If the estimated delivery result satisfies the information delivery condition, recommending the information to be recommended to the candidate user includes: If the candidate user belongs to the target category population, the information to be recommended is recommended to the terminal device used by the candidate user.

9. The method according to claim 6, characterized in that Outputting the estimated delivery result of the information to be recommended for the candidate user through the information recommendation estimation model includes: Determining the delivery score of the information to be recommended for the candidate user by using the information recommendation estimation model; If the estimated delivery result satisfies the information delivery condition, recommending the information to be recommended to the candidate user includes: When the delivery score exceeds a preset threshold, the information to be recommended is recommended to the terminal device used by the candidate user.

10. The method according to claim 1, characterized in that The determining of the exposure click feature corresponding to the candidate user according to the information recommendation status of the recommended information of the candidate user in the historical time period includes: Obtaining a first information recommendation status of the candidate user for a target information push position within a historical time period; the first information recommendation status includes a push status of at least one recommended information on the target information push position; the target information push position is the information push position corresponding to the information to be recommended; Determining the number of information exposures and the number of information clicks of the at least one recommended information in the target information push position according to the first information recommendation status; The number of information exposures and the number of information clicks of the candidate user on the target information push position are determined as the exposure-click features corresponding to the candidate user.

11. The method according to claim 1, wherein The determining of the exposure click feature corresponding to the candidate user according to the information recommendation status of the recommended information of the candidate user in the historical time period includes: Obtaining a second information recommendation status of the candidate user for target recommended information within a historical time period; the second information recommendation status includes a push status of the target recommended information on at least one information push position; the target recommended information is of the same type as the information to be recommended; Determining the number of information exposures and the number of information clicks of the target recommended information in at least one information push position according to the second information recommendation status; The number of information exposures and the number of information clicks of the target recommended information on at least one information push position are determined as the exposure-click features corresponding to the candidate user.

12. The method according to any one of claims 3 to 5, characterized in that The method further comprises: Acquire status information corresponding to the information to be recommended; the status information includes industry information, target placement information, and graphic content information corresponding to the information to be recommended; Acquire candidate information according to the status information corresponding to the information to be recommended, wherein the industry information corresponding to the candidate information is the same as the industry information corresponding to the information to be recommended, and the target placement information corresponding to the candidate information is the same as the target placement information corresponding to the information to be recommended; Obtaining a first feature vector corresponding to the information to be recommended based on the graphic content information corresponding to the information to be recommended; Obtaining a second feature vector corresponding to the candidate information according to the graphic content information corresponding to the candidate information; The historical recommended information corresponding to the to-be-recommended information is determined from the candidate information according to the cosine similarity between the first feature vector and the second feature vector.

13. An information recommendation device, characterized in that: The information recommendation device includes: An acquisition unit, configured to acquire user portrait features corresponding to candidate users for whom information is to be recommended; a determining unit, configured to determine an exposure-click feature corresponding to the candidate user based on an information recommendation status of recommended information of the candidate user within a historical time period, wherein the recommended information within the historical time period is of the same or similar type as the information to be recommended; the information recommendation status is a push status of the recommended information on an information push position; and the exposure-click feature includes a number of information exposures and a number of information clicks; The determining unit is further configured to determine an estimated exposure rate, an estimated click rate, and an estimated conversion rate of the information to be recommended for the candidate user based on the user portrait feature and the exposure-click feature of the candidate user; a processing unit, configured to determine an estimated delivery result of the information to be recommended for the candidate user based on the estimated exposure rate, the estimated click-through rate, and the estimated conversion rate of the information to be recommended for the candidate user; A sending unit is configured to recommend the information to be recommended to a terminal device used by the candidate user if the estimated delivery result satisfies an information delivery condition.

14. The device according to claim 13, characterized in that The determining unit, specifically the user: Inputting the user portrait features and the exposure-click features of the candidate user into a first model, and determining an estimated exposure rate of the to-be-recommended information for the candidate user through the first model; Inputting the user portrait features and the exposure-click features of the candidate user into a second model, and determining an estimated click-through rate of the information to be recommended for the candidate user through the second model; The user portrait features of the candidate user are input into a third model, and the estimated conversion rate of the information to be recommended for the candidate user is determined by the third model.

15. The device according to claim 14, characterized in that The device further comprises: a training unit; The acquisition unit is further configured to acquire demographic data and exposure demographic data corresponding to at least one historical recommendation information; the historical recommendation information and the information to be recommended are of the same information type; The determining unit is further configured to determine, based on the population targeting data and the exposed population data, unexposed population data corresponding to the historical recommendation information; determine the exposed population data as first positive sample training data, and determine the unexposed population data as first negative sample training data; wherein each first positive sample training data and each first negative sample training data includes user portrait features and exposure click features of a historical user, as well as actual exposure results of the historical recommendation information for the historical user; The training unit is configured to train the first model based on the first positive sample training data and the first negative sample training data.

16. The device according to claim 14, characterized in that The device further comprises: a training unit; The acquisition unit is further configured to acquire exposure population data and click population data corresponding to at least one historical recommendation information; the historical recommendation information and the information to be recommended are of the same information type; The determining unit is further configured to determine, based on the exposed population data and the click population data, exposed but not clicked population data corresponding to the historical recommendation information; determine the click population data as second positive sample training data, and determine the exposed but not clicked population data as second negative sample training data; wherein each second positive sample training data and each second negative sample training data includes user portrait features and exposure-click features of historical users, as well as actual click results of the historical recommendation information for the historical users; The training unit is used to train the second model based on the second positive sample training data and the second negative sample training data.

17. The device according to claim 14, characterized in that The device further comprises: a training unit; The acquisition unit is further configured to acquire click population data and conversion population data corresponding to at least one historical recommendation information; the historical recommendation information and the information to be recommended are of the same information type; The determining unit is further configured to determine, based on the click population data and the conversion population data, click non-conversion population data corresponding to the historical recommendation information; determine the conversion population data as third positive sample training data, and determine the click non-conversion population data as third negative sample training data; wherein each third positive sample training data and each third negative sample training data includes user portrait features of historical users, and actual conversion results of the historical recommendation information for the historical users; The training unit is used to train the third model based on the third positive sample training data and the third negative sample training data.

18. The device according to any one of claims 14 to 17, characterized in that The determination unit is specifically used to input the estimated exposure rate, estimated click rate and estimated conversion rate of the information to be recommended for candidate users into the information recommendation estimation model, and output the estimated delivery result of the information to be recommended for candidate users through the information recommendation estimation model.

19. The device according to claim 18, characterized in that The device further comprises: a training unit; The acquisition unit is further configured to acquire demographic targeting data corresponding to at least one historical recommendation information, wherein the historical recommendation information and the information to be recommended are of the same information type; the demographic targeting data includes user portrait features and exposure-click features of historical users, as well as actual delivery results of the historical recommendation information for the historical users; The acquisition unit is further configured to input the user portrait feature and the exposure-click feature into a first model, and obtain an estimated exposure rate of the historical recommendation information for the historical user through the first model; input the user portrait feature and the exposure-click feature into a second model, and obtain an estimated click rate of the historical recommendation information for the historical user through the second model; and input the user portrait feature into a third model, and obtain an estimated conversion rate of the historical recommendation information for the historical user through the third model; The determining unit is further configured to input the estimated exposure rate, estimated click rate, and estimated conversion rate of the historical recommendation information for the historical user into the information recommendation estimation model, and obtain the estimated delivery result of the historical recommendation information for the historical user through the information recommendation estimation model; The training unit is used to train the information recommendation prediction model according to a loss function, wherein the loss function is used to evaluate the similarity between the actual delivery result and the predicted delivery result of the historical recommendation information for the historical user.

20. The device according to claim 18, characterized in that The determining unit is specifically configured to classify the candidate users using the information recommendation estimation model; The sending unit is specifically configured to recommend the information to be recommended to a terminal device used by the candidate user if the candidate user belongs to a target category of people.

21. The device according to claim 18, characterized in that The determining unit is specifically configured to determine a delivery score of the information to be recommended for the candidate user through the information recommendation estimation model; The sending unit is specifically configured to recommend the information to be recommended to the terminal device used by the candidate user when the delivery score exceeds a preset threshold.

22. The device according to claim 13, characterized in that The acquisition unit is further configured to acquire a first information recommendation status of the candidate user for a target information push position within a historical time period; the first information recommendation status includes a push status of at least one recommended information on the target information push position; the target information push position is an information push position corresponding to the information to be recommended; The determination unit is specifically used to determine the number of information exposures and information clicks of the at least one recommended information on the target information push position based on the first information recommendation status; and determine the number of information exposures and information clicks of the candidate user on the target information push position as the exposure-click features corresponding to the candidate user.

23. The device according to claim 13, characterized in that The acquisition unit is further configured to acquire a second information recommendation status of the candidate user for the target recommended information within a historical time period; the second information recommendation status includes a push status of the target recommended information on at least one information push position; the target recommended information is of the same type as the information to be recommended; The determination unit is specifically used to determine the number of information exposures and the number of information clicks of the target recommended information on at least one information push position based on the second information recommendation status; and determine the number of information exposures and the number of information clicks of the target recommended information on at least one information push position as the exposure-click features corresponding to the candidate user.

24. The device according to any one of claims 15 to 17, characterized in that The acquisition unit is further configured to acquire status information corresponding to the information to be recommended; the status information includes industry information, target placement information, and graphic content information corresponding to the information to be recommended; The acquisition unit is further configured to acquire candidate information based on the status information corresponding to the information to be recommended, wherein the industry information corresponding to the candidate information is the same as the industry information corresponding to the information to be recommended, and the target placement information corresponding to the candidate information is the same as the target placement information corresponding to the information to be recommended; The determination unit is further used to obtain a first feature vector corresponding to the information to be recommended based on the graphic content information corresponding to the information to be recommended; obtain a second feature vector corresponding to the candidate information based on the graphic content information corresponding to the candidate information; and determine the historical recommendation information corresponding to the information to be recommended from the candidate information based on the cosine similarity between the first feature vector and the second feature vector.

25. A server, characterized in that: include: memories, transceivers, processors, and bus systems; Wherein, the memory is used to store programs; The processor is configured to execute the program in the memory to implement the method according to any one of claims 1 to 12; The bus system is used to connect the memory and the processor so that the memory and the processor can communicate with each other.

26. A computer-readable storage medium comprising instructions, which, when executed on a computer, cause the computer to perform the method according to any one of claims 1 to 12.

27. A computer program product, characterized in that The method comprises computer instructions, which are stored in a computer-readable storage medium; a processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the method according to any one of claims 1 to 12.

Images