Object recommendation model training method, related method, device, medium and equipment

Abstract

The embodiment of the invention discloses an object recommendation model training method, a related method, a device, a medium and equipment. An intersection object group between a first object group and a second object group of a promoter is obtained, sample object groups except the intersection object group are obtained from the first object group, then training data corresponding to the intersection object group and the sample object groups are obtained, a basic object recommendation model is trained according to the training data, and a first object recommendation model is obtained. And inputting the data corresponding to the to-be-screened object groups and the data corresponding to the target push information of the promotion platform into the first object recommendation model, and screening out a third object group in the to-be-screened object groups. Determining a fourth object group, a fifth object group and incremental training data corresponding to the fourth object group and the fifth object group in the third object group; and training the first object recommendation model according to the incremental training data to obtain a second object recommendation model. The model can accurately determine the object matched with the pushed information.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and in particular to an object recommendation model training method and related methods, devices, media and equipment. Background Art

[0002] With the development of network technology, a large number of Internet objects have emerged. Among these Internet objects, in order to generate better advertising revenue, it is necessary to find suitable target groups for some advertisements to achieve more accurate advertising delivery.

[0003] In the related art, one method for the target object group corresponding to the advertisement is to manually select the corresponding target object group, and then when the advertisement is delivered, the advertisement is delivered to the selected object group. Another method is to determine the target object group only by the click-through rate of the advertisement, and then deliver the advertisement to the target object group.

[0004] However, the above manual selection method is relatively fixed and cannot automatically cover more target groups, resulting in a single target group for advertising. It is also impossible to accurately determine whether an object is a target object only by the click-through rate. That is, the prior art cannot automatically and accurately select the target group of an advertisement. Summary of the invention

[0005] The embodiment of the present application provides an object recommendation model training method and related methods, devices, media and equipment. The trained object recommendation model can automatically and accurately determine the target object group to which the push information is adapted from the object group.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] A method for training an object recommendation model, comprising:

[0008] Obtaining an intersection object group between the first object group and the second object group of the promoter, and obtaining a sample object group other than the intersection object group in the first object group;

[0009] Acquire the first training data corresponding to the intersection object group, the second training data corresponding to the sample object group, and the historical push information data of the delivered historical push information;

[0010] The basic object recommendation model is trained according to the first training data, the second training data and the historical push information data to obtain a first object recommendation model;

[0011] Inputting the input object data corresponding to the to-be-screened object group and the target push information data of the target push information pushed by the promotion platform into the first object recommendation model, and screening out a third object group corresponding to the target push information from the to-be-screened object group;

[0012] Acquire a fourth target group selected by the promoter and a fifth target group not selected from the third target group;

[0013] Acquire first incremental training data corresponding to the fourth object group and second incremental training data corresponding to the fifth object group;

[0014] The first object recommendation model is trained according to the first incremental training data, the second incremental training data and the target push information data to obtain a second object recommendation model.

[0015] An object prediction method, comprising:

[0016] Get the input object group and target push information;

[0017] Acquire first target input object data and target push information data of target push information corresponding to the input object group;

[0018] The first target input object data and the target push information data are input into the second object recommendation model trained by the object recommendation model training method provided in the embodiment of the present application to predict the first target object corresponding to the target push information.

[0019] An object recommendation model training device, comprising:

[0020] A first acquisition module is used to acquire an intersection object group between the first object group and the second object group of the promoter, and acquire a sample object group other than the intersection object group in the first object group;

[0021] A second acquisition module is used to acquire the first training data corresponding to the intersection object group, the second training data corresponding to the sample object group, and the historical push information data of the delivered historical push information;

[0022] A first training module, used to train the basic object recommendation model according to the first training data, the second training data and the historical push information data to obtain a first object recommendation model;

[0023] A screening module, used to input the input object data corresponding to the to-be-screened object group and the target push information data of the target push information pushed by the promotion platform into the first object recommendation model, and screen out a third object group corresponding to the target push information from the to-be-screened object group;

[0024] A third acquisition module is used to acquire a fourth target group selected by the promoter and a fifth target group not selected from the third target group;

[0025] A fourth acquisition module, used to acquire first incremental training data corresponding to a fourth object group and second incremental training data corresponding to a fifth object group;

[0026] The second training module is used to train the first object recommendation model according to the first incremental training data, the second incremental training data and the target push information data to obtain the second object recommendation model.

[0027] In some embodiments, the second acquisition module further includes a first sub-acquisition module and a second sub-acquisition module.

[0028] A first sub-acquisition module, used to acquire first object data of each object in the intersection object group, and first push information data of push information browsed by each object in the intersection object group;

[0029] Generate first training data corresponding to the intersection object group according to the first object data and the first push information data of each object in the intersection object group;

[0030] A second sub-acquisition module, used to acquire second object data of each object in the sample object group, and second push information data of push information browsed by each object in the sample object group;

[0031] Second training data corresponding to the sample object group is generated according to the second object data and the second push information data of each object in the sample object group.

[0032] In some implementations, the second training module includes a first sub-feature acquisition module, a second sub-feature acquisition module, and a sub-training module.

[0033] A first sub-feature acquisition module, used to input the first training data into the object tower, output the first extracted feature, and input the second training data into the object tower, output the second extracted feature;

[0034] The second sub-feature acquisition module is used to input the historical push information data into the item tower and output the third extracted feature;

[0035] A sub-training module, used for calculating the similarity between the first extracted feature and the third extracted feature to obtain a first calculation result;

[0036] Calculate the similarity between the second extracted feature and the third extracted feature to obtain a second calculation result;

[0037] The basic object recommendation model is trained according to the first calculation result and the second calculation result to obtain a first object recommendation model.

[0038] In some embodiments, the first sub-feature acquisition module is used to input the first object data in the first training data into the first sub-object tower and output the first feature;

[0039] Inputting the first push information data in the first training data into at least one second sub-object tower, and outputting a second feature;

[0040] The first feature and the second feature are input into a multilayer perceptron, and a first extracted feature is output.

[0041] In some embodiments, the first sub-feature acquisition module is used to input the second object data in the second training data into the first sub-object tower and output the third feature;

[0042] Inputting the second push information data in the second training data into at least one second sub-object tower, and outputting a fourth feature;

[0043] The third feature and the fourth feature are input into the multilayer perceptron, and the second extracted feature is output.

[0044] In some embodiments, the sub-training module is used to input the first extracted feature into the output layer and input the third extracted feature into the output layer;

[0045] Calculate the similarity between the first extracted feature and the third extracted feature at the output layer according to the noise contrast estimation loss function to obtain a first calculation result;

[0046] Inputting the second extracted feature into the output layer, and inputting the third extracted feature into the output layer;

[0047] The similarity between the second extracted feature and the third extracted feature is calculated at the output layer according to the noise contrast estimation loss function to obtain a second calculation result.

[0048] In some embodiments, the second sub-feature acquisition module is used to determine, in the push information data corresponding to at least one second sub-object tower, selected push information data that is at least partially identical to the historical push information data before inputting the historical push information data into the object tower and outputting the third extracted feature;

[0049] The vector parameter represented by the selected push information data in the second sub-object tower is determined as the target vector parameter in the item tower, and the corresponding selected push information data in the item tower is set as the target vector parameter.

[0050] In some implementations, the first acquisition module includes a first sub-encryption module, a second sub-encryption module, and a sub-decryption module.

[0051] A first sub-encryption module, used to determine a first encryption result corresponding to first group information of a first object group;

[0052] A second sub-encryption module, used to determine a second encryption result corresponding to the second group information of the second target group of the promoter;

[0053] A sub-decryption module, used for performing intersection acquisition processing on the first encryption result and the second encryption result to obtain an intersection result;

[0054] The intersection result is decrypted to obtain an intersection object group between the first object group and the second object group.

[0055] In some implementations, the first sub-encryption module is used to encrypt first group information of the first object group to obtain first encrypted information;

[0056] Sending the first encrypted information to the promoter;

[0057] The second encrypted information returned by the promoter after encrypting the first encrypted information is received, and the second encrypted information is determined as the first encryption result.

[0058] In some embodiments, the second sub-encryption module is used to obtain third encrypted information after the promoter encrypts the second group information of the second target group;

[0059] The third encrypted information is encrypted to obtain fourth encrypted information, and the fourth encrypted information is determined as the second encryption result.

[0060] In some embodiments, the third acquisition module is used to send a third target group to the promoter;

[0061] receiving a fourth target group selected by the promoter from the third target group;

[0062] A fifth target group not selected by the promoter from the third target group is received.

[0063] In some embodiments, the fourth acquisition module is used to acquire third object data of each object in the fourth object group, and third push information data of push information browsed by each object in the fourth object group;

[0064] generating first incremental training data corresponding to the fourth object group according to the third object data and the third push information data of each object in the fourth object group;

[0065] Acquire fourth object data of each object in the fifth object group, and fourth push information data of push information browsed by each object in the fifth object group;

[0066] Second incremental training data corresponding to the sample object group is generated according to the fourth object data and the fourth push information data of each object in the fifth object group.

[0067] In some embodiments, the second training module is used to input the first incremental training data into the object tower and output the fourth extracted feature, input the second incremental training data into the object tower and output the fifth extracted feature, and input the third incremental training data into the object tower and output the sixth extracted feature;

[0068] Input the target push information data into the item tower and output the seventh extraction feature;

[0069] Calculate the similarity between the fourth extracted feature and the seventh extracted feature to obtain a third calculation result;

[0070] Calculate the similarity between the fifth extracted feature and the seventh extracted feature to obtain a fourth calculation result;

[0071] Calculate the similarity between the sixth extracted feature and the seventh extracted feature to obtain a fifth calculation result;

[0072] The first object recommendation model is trained according to the third calculation result, the fourth calculation result and the fifth calculation result to obtain a second object recommendation model.

[0073] An object prediction device, characterized by comprising:

[0074] An input acquisition module is used to obtain input object groups and target push information;

[0075] A data acquisition module, used to acquire first target input object data and target push information data of target push information corresponding to the input object group;

[0076] The prediction module is used to input the first target input object data and the target push information data into the second object recommendation model trained by the object recommendation model training method provided in the embodiment of the present application, and predict the first target object corresponding to the target push information.

[0077] In some implementations, the prediction module is further used to obtain the total number of objects requiring target push information and the importance of the target push information;

[0078] determining the number of objects in the input object population;

[0079] Determine the ratio of the number of first target objects to the number of objects;

[0080] Generate constraints based on total number of objects, importance, number of objects, and occupancy ratio;

[0081] Based on the allocation algorithm and constraints, the target push information is delivered to the input object group.

[0082] A computer-readable storage medium stores a plurality of instructions suitable for loading by a processor to execute the steps in the above-mentioned object recommendation model training method or the steps in the object prediction method.

[0083] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps in the above-mentioned object recommendation model training method or the steps in the object prediction method when executing the computer program.

[0084] A computer program product or a computer program, the computer program product or the computer program comprising computer instructions, the computer instructions being stored in a storage medium. A processor of a computer device reads the computer instructions from the storage medium, and the processor executes the computer instructions to implement the steps in the above-mentioned object recommendation model training method or the steps in the object prediction method.

[0085] In an embodiment of the present application, an intersection object group between a first object group and a second object group of a promoter is obtained, and a sample object group other than the intersection object group is obtained in the first object group; first training data corresponding to the intersection object group, second training data corresponding to the sample object group, and historical push information data of historical push information that has been delivered are obtained; a basic object recommendation model is trained according to the first training data, the second training data, and the historical push information data to obtain a first object recommendation model; input object data corresponding to the object group to be screened and target push information data of the target push information pushed by the promotion platform are input into the first object recommendation model, and a third object group corresponding to the target push information is screened out from the object group to be screened; a fourth object group selected by the promoter and an unselected fifth object group are obtained in the third object group; first incremental training data corresponding to the fourth object group and second incremental training data corresponding to the fifth object group are obtained; the first object recommendation model is trained according to the first incremental training data, the second incremental training data, and the target push information data to obtain a second object recommendation model. In this way, through the interaction between the promoter and the service provider, the common intersection object group and the sample object group without intersection are determined, and the basic object recommendation is preliminarily trained in combination with the historical push information that has been delivered to obtain a first object recommendation model. The first object recommendation model is used to preliminarily predict the third object group corresponding to the target push information, and the third object group is divided into a selected fourth object group and an unselected fifth object group according to the needs of the promoter. The first object recommendation model is trained for the second time to obtain a trained second object recommendation model. Compared with the scheme of manually circling or selecting the target object group by single click rate, the trained second object recommendation model of the present application is more suitable for the promoter's target push information, and can automatically and accurately determine the target object group that is compatible with the target push information among the object groups.

[0086] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained by the structures specifically pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0088] Figure 1 is a schematic diagram of a scene for object prediction provided by an embodiment of the present application;

[0089] Figure 2 It is a flowchart of the object recommendation model training method provided in an embodiment of the present application;

[0090] Figure 3 is a schematic diagram of a process for obtaining an intersection object group provided in an embodiment of the present application;

[0091] Figure 4 is a schematic diagram of a scenario for obtaining an intersection object group provided in an embodiment of the present application;

[0092] Figure 5 is a schematic diagram of the structure of the object recommendation model provided in the embodiment of the present application;

[0093] Figure 6 It is a flowchart of the basic object recommendation model training provided in the embodiment of the present application;

[0094] Figure 7 is a flowchart of the first object recommendation model training provided in an embodiment of the present application;

[0095] Figure 8 is another flowchart of the object recommendation model training method provided in an embodiment of the present application;

[0096] Fig. 9 is a schematic diagram of a scenario corresponding to the object recommendation model provided in an embodiment of the present application;

[0097] Fig.10 is a flow chart of an object prediction method provided in an embodiment of the present application;

[0098] Fig.11 It is a schematic diagram of a scenario of applying the allocation algorithm provided in an embodiment of the present application;

[0099] Fig.12 is a schematic diagram of the structure of a third object recommendation model provided in an embodiment of the present application;

[0100] Fig.13 is a schematic diagram of the structure of an object recommendation model device provided in an embodiment of the present application;

[0101] Fig.14 is a schematic diagram of the structure of an object prediction device provided in an embodiment of the present application;

[0102] Fig.15 It is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0103] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0104] It should be noted that in some processes described in the specification, claims and the above-mentioned drawings, multiple steps appearing in a specific order are included, but it should be clearly understood that these steps may not be executed in the order in which they appear in this document or may be executed in parallel. The step numbers are only used to distinguish different steps, and the numbers themselves do not represent any execution order. In addition, descriptions such as "first", "second" or "target" in this document are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0105] It is understandable that in the specific implementation of this application, it involves obtaining relevant data such as user information. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0106] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations:

[0107] Private Set Intersection (PSI) refers to the process of obtaining the intersection of the data held by both parties without revealing any additional information. Here, additional information refers to any information other than the intersection of the data held by both parties.

[0108] Bloom Filter, the main principle of Bloom filter is to use a set of hash functions to map elements to index positions in a set of bit arrays. When you want to check whether an element is in a set, hash the element and then check whether the value of the bit array corresponding to the hash value is 1. If the values ​​of the bit arrays corresponding to the hash values ​​are all 1, then the element may be in the set, otherwise the element is not in the set. It can quickly check whether an element belongs to a set. It can quickly determine whether an element is in a large set, and the judgment speed is very fast and does not take up too much memory space.

[0109] DH (Diffie-Hellman) key exchange algorithm enables the communicating parties to securely exchange keys in an insecure channel for encrypting subsequent communication messages.

[0110] Embedding, in machine learning and natural language processing, refers to the process of mapping high-dimensional data (such as text, images, audio) to a low-dimensional space. The embedding vector is usually a vector of real numbers, which represents the input data as a point in a continuous numerical space.

[0111] Multilayer Perceptron (MLP) is a feed-forward artificial neural network model that maps multiple input data sets to a single output data set.

[0112] The fully connected layer (Full Connection) is a layer in which each node is connected to all nodes in the previous layer to integrate the features extracted previously.

[0113] Noise Contrastive Estimation (NCE) is a new statistical model estimation method that can be used to solve complex computational problems of neural networks, and is therefore widely used in image processing and natural language processing. For example, a multi-classification problem can be converted into a binary classification problem, and the characteristics of the data can be mined by learning the difference between the data distribution and the noise distribution.

[0114] Demand-Side Platform (DSP) is a system and an online advertising platform in the Internet advertising industry. It serves advertisers and helps them place advertisements on the Internet or mobile Internet. The DSP allows advertisers to follow a unified bidding and feedback method more easily and conveniently, and purchase high-quality advertising inventory in real time at a reasonable price for online advertisements located on multiple advertising trading platforms. The DSP allows advertisers to manage one or more Ad Exchange accounts through a unified interface, and even the DSP can help advertisers manage Ad Exchange accounts, providing a full range of services.

[0115] See also Figure 1 , Figure 1 It is a schematic diagram of a scene for object prediction provided in an embodiment of the present application.

[0116] The entire scenario includes the service provider, promoter, terminal and object, where the object can be understood as the object that uses the service corresponding to the service provider.

[0117] In some application scenarios, the promoter needs to place corresponding push information, such as advertisements and videos, on the service provider's platform for the purpose of publicity and marketing. The service provider can place push information to the objects using its services.

[0118] When the subject uses the service provided by the service provider through the terminal, the service provider can push the push information to the terminal, thereby completing the delivery of the push information to the subject through the client provided by the service provider on the terminal. For example, when the subject uses instant messaging software, video software, music software, and reading software, if the services of these software are provided by the service provider, the service provider can push the push information to the client provided by these software through these software, so that the subject can see the corresponding push information when using the client provided by the terminal.

[0119] It is understandable that in the process of delivering push information to multiple objects, not every object is the target object required by the promoter. For example, although some object groups browse push information, these object groups are not the target groups required by the promoter. For example, the push information corresponding to the promoter is mainly delivered to objects in the age group of 20 to 30 years old, and objects in the age group of 20 to 30 years old are the main sales group of the corresponding products or services of the push information, but some of the object groups that browse push information are in the age group of 60 to 70 years old. Therefore, the object group in the age group of 60 to 70 years old is not the target object group required by the promoter.

[0120] In conventional technologies, some of them select the target object group of the push information through manual screening from a large number of object groups, and some of them simply determine the target object group of the push information through the click-through rate of the push information or the exposure rate of the push information. The manual screening method will result in the omission of some potential target objects in the target object group obtained, and there will be some non-target users who are not related to the push information in the target object group obtained through click-through rate or exposure rate. Therefore, how the service provider can accurately select the target object group that the promoter's push information is adapted to from the service object group is a technical problem that needs to be solved.

[0121] In the embodiment of the present application, from the perspective of the service provider, an object recommendation model is set up to automatically and accurately screen out target objects that are suitable for the target push information of the promoter from the object group served by the service provider.

[0122] Among them, for the training of the object recommendation model, the embodiment of the present application is to conduct training through the interaction between the service party and the promoter, and then perform training based on the training data after the interaction, so that the object recommendation model can better meet the promotion policy's requirements for obtaining the target objects for target push information.

[0123] For example, before delivering the target push information, the service provider can obtain the intersection object group between the service provider and the promoter, and then the service provider can obtain the sample object group of the local first object group excluding the intersection object group. Then, the basic object recommendation model is trained using the first training data corresponding to the intersection object group and the second training data corresponding to the sample object group, as well as the historical push data of the historical push information delivered historically, thereby obtaining the first object recommendation model.

[0124] It can be understood that the first object recommendation model is a model that preliminarily meets the target push information needs of the promoter, and the first object recommendation model can preliminarily screen out the third object group corresponding to the target push information from the to-be-screened population. The third object group can be considered as the object group that the first object recommendation model believes matches the target push information in the to-be-screened object group. The service provider can send the third object group to the promoter.

[0125] The promoter can select a part of the third target group according to its own demand standards and determine it as the fourth target group associated with the target push information. The target group in the third target group that is not selected by the push party is determined as the fifth target group that is not associated with the target push information.

[0126] The service provider can obtain the fourth object group and the fifth object group, and then obtain the first incremental training data corresponding to the fourth object group and the second incremental training data corresponding to the fifth object group, obtain the target push information data of the target push information, and then train the first object recommendation model according to the first incremental training data, the second incremental training data and the target push information data, so as to obtain the second object recommendation model.

[0127] It is understandable that, since the first object recommendation model is trained by the training data of the target group selected by the promoter and the target push information data of the target push information, the obtained second object recommendation model is more in line with the audience needs of the promoter's target push information, and can better select the target object group of the target push information from the object group served by the service provider. Thus, the target object group adapted to the target push information is automatically and accurately determined from the object group.

[0128] For a more detailed description of the object recommendation model training method provided in the embodiment of the present application, please continue to refer to Figure 2 , Figure 2 : is a flowchart of the object recommendation model training method provided by the embodiment of the present application. The following object recommendation model training method is considered from the perspective of the service provider, and the object recommendation model training method may include the following steps:

[0129] In step 110, an intersection object group between the first object group and the second object group of the promoter is obtained, and a sample object group other than the intersection object group is obtained from the first object group.

[0130] It is understandable that before training the basic object recommendation model, it is also necessary to obtain the training samples corresponding to the basic object recommendation model, where the training samples include positive samples and negative samples. Considering that the object recommendation model serves the promoter, the intersection object group between the local first object group and the promoter's second object group can be obtained. Specifically, the intersection object group can be obtained by intersecting the privacy sets, thereby ensuring that when the intersection object group is obtained, no user privacy will be leaked between the service provider and the promoter. Since the intersection object group is also the object group saved by the promoter, the data corresponding to the intersection object group can be understood as the positive sample in the training sample of the basic object recommendation model.

[0131] After obtaining the intersection object group, a sample object group other than the intersection object group can be obtained from the first object group. For example, in the group other than the intersection object group in the first object group, a portion of the sample object group can be screened out by random screening, and the data corresponding to the sample object group can be understood as a negative sample in the training sample of the object recommendation model.

[0132] Please also read Figure 3 , Figure 3 It is a flowchart of obtaining an intersection object group provided in an embodiment of the present application.

[0133] In some implementations, obtaining an intersection object group between the first object group and the second object group of the promoter includes:

[0134] In step 201, a first encryption result corresponding to first group information of a first object group is determined.

[0135] It is understandable that in order to protect the privacy information of the objects stored locally from being leaked between the service provider and the promoter, the intersection object group between the two can be obtained by intersecting the privacy sets. For example, the service provider can encrypt the first group information of the first object group to protect the privacy information of the object.

[0136] In some implementations, determining a first encryption result corresponding to first group information of a first object group includes:

[0137] (1.1) encrypting first group information of a first target group to obtain first encrypted information;

[0138] (1.2) sending the first encrypted information to the promoter;

[0139] (1.3) Receive the second encrypted information returned by the promoter after encrypting the first encrypted information, and determine the second encrypted information as the first encryption result.

[0140] Please also read Figure 4 , Figure 4 It is a schematic diagram of a scenario for obtaining an intersection object group provided in an embodiment of the present application.

[0141] The service provider may encrypt the first group information of the first target group, for example, the service provider may establish a private key to initially encrypt the first group information of the first target group, thereby obtaining first encrypted information. The service provider then sends the first encrypted information to the promoter.

[0142] After receiving the first encrypted information, the promoter may encrypt the first encrypted information, for example, by using the public key between the service provider and the promoter to encrypt the first encrypted information, thereby obtaining the second encrypted information. The promoter then sends the second encrypted information to the service provider.

[0143] After receiving the second encrypted information, the service provider may determine the second encrypted information as the first encryption result.

[0144] In some embodiments, before the service provider sends the first encrypted information to the pusher, the service provider may establish a data connection with the pusher in advance, and then the service provider may send an intersection object acquisition request to the promoter. After the promoter agrees to the intersection object acquisition request, the service provider may send the first encrypted information to the promoter.

[0145] It can be understood that the first group information is encrypted twice, thereby preventing the privacy information of the objects in the first group information from being leaked.

[0146] In step 202, a second encryption result corresponding to second group information of a second target group of the promoter is determined.

[0147] It is understandable that the promoter can also encrypt the stored second group information of the second object group to protect the privacy information of the object. The service provider can encrypt the encrypted second group information again to obtain a second encryption result.

[0148] In some implementations, determining a second encryption result corresponding to the second group information of the second target group of the promoter includes:

[0149] (2.1) obtaining third encrypted information after the promoter encrypts the second group information of the second target group;

[0150] (2.2) The third encrypted information is encrypted to obtain fourth encrypted information, and the fourth encrypted information is determined as the second encryption result.

[0151] like Figure 4 As shown, the promoter can encrypt the second group information of the second target group to obtain the third encrypted information. For example, the promoter can initially encrypt the second group information of the second target group by establishing a private key to obtain the third encrypted information. The promoter then sends the third encrypted information to the service provider.

[0152] The service provider may obtain the third encrypted information sent by the promoter, and then encrypt the third encrypted information to obtain the fourth encrypted information, and determine the fourth encrypted information as the second encryption result. For example, the service provider may encrypt the third encrypted information again using the public key between the service provider and the promoter to obtain the fourth encrypted information, and finally determine the fourth encrypted information as the second encryption result.

[0153] In step 203, intersection acquisition processing is performed on the first encryption result and the second encryption result to obtain an intersection result.

[0154] After the first encryption result and the second encryption result are obtained, an intersection acquisition process may be performed on the first encryption result and the encryption result to obtain an intersection result.

[0155] Understandably, if you want to determine whether an element is in a set, you usually think of saving all the elements and then confirming by comparison. However, as the number of elements in the set increases, the storage space required becomes larger and the retrieval speed becomes slower.

[0156] If there are multiple promoters, simultaneously performing intersection acquisition processing on the encryption results corresponding to the multiple promoters and the encryption results of the service provider will inevitably lead to an increase in the required memory capacity, thereby affecting the computing performance of the service provider.

[0157] In order to avoid this technical problem, in the stage of obtaining the intersection of the first encryption result and the second encryption result, a Bloom filter is used in the embodiment of the present application to realize the intersection of the first encryption result and the second encryption result, thereby reducing the consumption of the memory capacity of the service party and improving the efficiency of the intersection of the first encryption result and the second encryption result.

[0158] Specifically, the first encryption result may be used to generate a data structure corresponding to a Bloom filter. For example, for each element in the first encryption result, multiple hash values ​​are generated through multiple hash functions, and then the target array position corresponding to each hash value is determined in a preset bit array, and the values ​​of these target array positions are set to 1.

[0159] Then, multiple hash operations are performed on the elements to be queried in the second encryption result, such as using multiple hash functions to perform hash calculations on the elements to be queried, thereby obtaining multiple hash values, and then determining the array position corresponding to each hash value in the data structure. If the value of the array position corresponding to each hash value in the data structure is 1, it means that the element to be queried exists in the first encryption result. The above elements can be understood as the object information of each object.

[0160] Through the above method, the intersection acquisition process of the first encryption result and the second encryption result is completed to obtain the intersection result.

[0161] In step 204, the intersection result is decrypted to obtain an intersection object group between the first object group and the second object group.

[0162] In some implementations, after obtaining the intersection result, a DH key exchange algorithm may be used to determine a target key corresponding to the intersection result, and the intersection result may be decrypted using the target key to obtain an intersection object group between the first object group and the second object group.

[0163] For example, the service provider can use its own private key and the public key between the service provider and the recommender to calculate the target key corresponding to the intersection result through the DH key exchange algorithm, and then decrypt the intersection result through the target key to obtain the object group information in the intersection result, and finally determine the intersection object group based on the object group information in the intersection result.

[0164] In step 120, first training data corresponding to the intersection object group, second training data corresponding to the sample object group, and historical push information data of delivered historical push information are obtained.

[0165] After obtaining the intersection object group, data corresponding to each object in the intersection object group can be obtained, and then first training data can be generated based on the data corresponding to each object in the intersection object group. After obtaining the sample object group, data corresponding to each object in the sample object group can be obtained, and then second training data can be generated based on the data corresponding to each object in the sample object group.

[0166] In some implementations, obtaining first training data corresponding to the intersection object group and second training data corresponding to the sample object group includes:

[0167] (1.1) obtaining the first object data of each object in the intersection object group and the first push information data of the push information browsed by each object in the intersection object group;

[0168] (1.2) generating first training data corresponding to the intersection object group according to the first object data and the first push information data of each object in the intersection object group;

[0169] (1.3) obtaining second object data of each object in the sample object group, and second push information data of push information browsed by each object in the sample object group;

[0170] (1.4) Generate second training data corresponding to the sample object group based on the second object data and the second push information data of each object in the sample object group.

[0171] The first object data of each object in the intersection object group includes basic attributes of the object, such as age, region, education level, preference, etc. These basic attributes of the object are obtained after authorization by the object, or are basic attributes actively input by the object when using the service of the service provider.

[0172] The first push information data of the push information browsed by each object in the intersection object group may also include: a schedule identifier of the push information, a push information identifier, an industry category identifier, and the like.

[0173] After the first object data and the first push information data are acquired, the two may be determined as first training data. The first training data may be understood as positive sample data corresponding to the basic object recommendation model during training.

[0174] The second object data of each object in the sample object group includes basic attributes of the object, such as age, region, education level, preference, etc. These basic attributes of the object are obtained after authorization by the object, or are basic attributes actively input by the object when using the service of the service provider.

[0175] The second push information data of the push information browsed by each object in the sample object group may also include: a schedule identifier of the push information, a push information identifier, an industry category identifier, and the like.

[0176] After the second object data and the second push information data are acquired, the two can be determined as second training data. The second training data can be understood as negative sample data corresponding to the basic object recommendation model during training.

[0177] In some implementations, it is also necessary to obtain historical push information data of historical push information that has been delivered. The delivered historical push information may be push information within a preset historical time period before the current time, and the historical push information may also be push information that the promoter has delivered to the service provider. The historical push information data may include: a schedule identifier of the push information, a push information identifier, and an industry category identifier, etc.

[0178] In step 130, the basic object recommendation model is trained according to the first training data, the second training data and the historical push information data to obtain a first object recommendation model.

[0179] The object recommendation model provided in the embodiment of the present application can be understood as a dual-tower model, which includes an object tower (user) and an item tower (item), wherein the object tower is used to input object-related data to extract object-related features, and the item tower is used to input item-related data to extract item-related features, and finally the object-related features and item-related features are compared for similarity and the results are output. Since the entire data processing flow is like two towers, it is called a dual-tower model.

[0180] Please also read Figure 5 , Figure 5 is a schematic diagram of the structure of the object recommendation model provided in the embodiment of the present application, such as Figure 5 As shown, the object recommendation model provided by the embodiment of the present application includes an object tower, an item tower, a multi-layer perceptron, and an output layer. The output of the object tower is connected to the input of the multi-layer perceptron, and the output of the multi-layer perceptron is connected to the input of the output layer. The output of the item tower is connected to the input of the output layer.

[0181] The first training data is used to input the object tower, the second training data is used to input the object tower, and the historical push information data of the historical push information that has been put into the item tower is input.

[0182] Please also read Figure 6 , Figure 6 It is a flowchart of the basic object recommendation model training provided in an embodiment of the present application.

[0183] In some implementations, training a basic object recommendation model according to the first training data, the second training data, and the historical push information data to obtain a first object recommendation model includes:

[0184] 301. Input the first training data into the object tower, output the first extracted features, and input the second training data into the object tower, output the second extracted features.

[0185] like Figure 5 As shown, the object tower includes multiple sub-object towers, specifically, the object tower includes a first sub-object tower and at least one second sub-object tower. Each sub-object tower is correspondingly provided with an embedding layer. When each sub-object tower inputs data, the embedding layer can extract the features of the data, thereby outputting vector features related to the input data.

[0186] From the above, it can be seen that the first training data includes first object data and first push information data, and the first object data and the first push information data are respectively input into different sub-object towers in the object tower, for example, the first object data is input into the first sub-object tower, and the first push information data is input into at least one second sub-object tower.

[0187] The second training data includes second object data and second push information data. The second object data and the second push information data are respectively input into different sub-object towers in the object tower. For example, the second object data is input into the first sub-object tower, and the second push information data is input into at least one second sub-object tower.

[0188] That is, in the embodiment of the present application, the object tower can be used to input object data of the object, such as age, region, education level and preference, etc. The object tower can also be used to input push information data, such as the schedule identification of the push information, the push information identification and the industry category identification, etc.

[0189] In the traditional dual-tower model, only some data related to the object can be entered in the object tower, namely the object's age, gender, and region. The relevant data of the push information, such as the schedule identifier, push information identifier, and industry category identifier, can only be entered into the item tower. This results in less feature interaction between the item tower and the object tower, and ultimately the trained model cannot accurately predict the objects related to the push information.

[0190] In the embodiment of the present application, considering the above technical problems of the traditional double-tower model, a first sub-object tower and at least one second sub-object tower are set in the object tower, the first sub-object tower is used to input the object data of the object, and the at least one second sub-object tower is used to input the push information data of the push information, thereby increasing the interaction between the object features and the push information features. That is, the feature interaction dimension between the object features and the item features is increased, which is conducive to improving the prediction accuracy of the subsequently trained object recommendation model for the object adapted to the push information.

[0191] In some implementations, inputting first training data into an object tower and outputting first extracted features comprises:

[0192] (1.1) inputting the first object data in the first training data into the first sub-object tower and outputting the first feature;

[0193] (1.2) inputting the first push information data in the first training data into at least one second sub-object tower, and outputting a second feature;

[0194] (1.3) The first feature and the second feature are input into a multilayer perceptron, and the first extracted feature is output.

[0195] like Figure 5As shown, the first sub-object tower can be any one of the multiple sub-object towers of the object tower. The first object data in the first training data is input into the first sub-object tower, and the features of the first object data are extracted through the embedding layer of the first sub-object tower, so as to output the first feature.

[0196] Specifically, data such as the age, region, education level, and preference of each object in the intersection object group may be input into the first object tower, thereby outputting the first feature corresponding to the first object data.

[0197] The second sub-object tower can be a sub-object tower in the object tower except the first sub-object tower. The first push information data is input into at least one second sub-object tower, and the features of the first push information data are extracted through the embedding layer of the second sub-object tower to obtain the second feature.

[0198] Specifically, data such as the schedule identifier, push information identifier, and industry category identifier of the push information browsed by each object in the intersection object group may be input into at least one second sub-object tower, thereby outputting the second feature corresponding to the first push information.

[0199] It should be noted that the second sub-object tower can be one, two or more. When the object has browsed multiple push information, the number of second sub-object towers can also be multiple, and then the push data of each push information browsed by the object is input into different second sub-object towers respectively.

[0200] After obtaining the first feature and the second feature, the first feature and the second feature are input into a multilayer perceptron, and the first extracted feature is output. The multilayer perceptron includes multiple fully connected layers, and the output of the previous fully connected layer is input into the next fully connected layer. The input of the first fully connected layer in the multilayer perceptron is connected to the output of the object tower, and the output of the last fully connected layer of the multilayer perceptron is the final output of the multilayer perceptron.

[0201] In some implementations, inputting the second training data into the object tower and outputting the second extracted features comprises:

[0202] (2.1) inputting the second object data in the second training data into the first sub-object tower, and outputting the third feature;

[0203] (2.2) inputting the second push information data in the second training data into at least one second sub-object tower, and outputting a fourth feature;

[0204] (2.3) The third feature and the fourth feature are input into the multilayer perceptron, and the second extracted feature is output.

[0205] The first sub-object tower may be any one of the multiple sub-object towers of the object tower. The second object data in the second training data is input into the first sub-object tower, and the features of the second object data are extracted through the embedding layer of the first sub-object tower, thereby outputting the first features.

[0206] Specifically, data such as the age, region, education level, and preference of each object in the sample object group may be input into the first object tower, thereby outputting the third feature corresponding to the second object data.

[0207] The second sub-object tower can be a sub-object tower in the object tower except the first sub-object tower. The second push information data is input into at least one second sub-object tower, and the features of the second push information data are extracted through the embedding layer of the second sub-object tower to obtain the fourth feature.

[0208] Specifically, data such as the schedule identifier, push information identifier, and industry category identifier of the push information browsed by each object in the sample object group can be input into at least one second sub-object tower, so as to output the fourth feature corresponding to the second push information.

[0209] After obtaining the third feature and the fourth feature, the third feature and the fourth feature are input into a multilayer perceptron, and the second extracted feature is output.

[0210] 302. Input the historical push information data into the item tower and output the third extracted feature.

[0211] In some embodiments, after the historical push information data is acquired, the historical push information data is input into an item tower, and the item tower includes an embedding layer, in which the historical push information data is feature extracted through the embedding layer to obtain a third extracted feature.

[0212] Specifically, data such as the schedule identifier, the push information identifier, and the industry category identifier in the historical push information data can be input into the item tower to obtain the third extracted feature.

[0213] In some embodiments, before inputting the historical push information data into the item tower and outputting the third extracted feature, the process includes:

[0214] (3.1) determining selected push information data that is at least partially identical to historical push information data from the push information data corresponding to at least one second sub-object tower;

[0215] (3.2) The vector parameter represented by the selected push information data in the second sub-object tower is determined as the target vector parameter in the item tower, and the corresponding selected push information data in the item tower is set as the target vector parameter.

[0216] It is understandable that in order to strengthen the feature interaction between the object tower and the item tower, thereby improving the accuracy of the trained object recommendation model in predicting the object that the push information is adapted to in actual applications, the parameters between the second sub-object tower and the item tower can be shared, such as vector parameters. The second sub-object tower and the item tower both have an embedding layer, and at least part of the vector parameters of the embedding layer of the second sub-object tower can be shared with the vector parameters of the item tower.

[0217] For example, the push information data corresponding to the second sub-object tower includes a scheduling identifier, a push information identifier, and an industry category identifier, and the historical push information data corresponding to the item tower also includes a scheduling identifier, a push information identifier, and an industry category identifier.

[0218] The category identifier in the push information data corresponding to the second sub-object tower is: category (425234), and the category identifier in the historical push information data corresponding to the item tower is: category (425234). This means that the category identifier of the push information data corresponding to the second sub-object tower is the same as the category identifier in the historical push information data corresponding to the item tower. Then category (425234) can be used as the selected push information data.

[0219] Then, the second sub-object tower corresponding to the selected push information data is determined. In the second sub-object tower corresponding to the selected push information data, if the vector parameter corresponding to the category identification of the push information data corresponding to the second sub-object tower is the vector [0.234, 0.434], the vector parameter is determined as the target vector parameter.

[0220] Then the vector parameter corresponding to the category identification in the historical push information data corresponding to the item tower is the target vector parameter, which is also the vector [0.234, 0.434].

[0221] That is to say, before the item tower extracts features from the historical push information data, the vector parameters corresponding to the embedding layer of the item tower can be set first, thereby increasing the interaction between the object tower and the item tower, which is more conducive to the prediction accuracy of the trained object recommendation model.

[0222] After the embedding layer target vector parameters of the item tower are set, the historical push information data can be input into the item tower with the target vector parameters set, and features are extracted through the embedding layer of the item tower, thereby outputting the third extracted features.

[0223] 303. Perform similarity calculation on the first extracted feature and the third extracted feature to obtain a first calculation result.

[0224] After obtaining the first extracted features and the third extracted features, the first extracted features and the third extracted features may be input into an output layer to calculate the similarity between the two, thereby obtaining a first calculation result.

[0225] In some implementations, performing similarity calculation on the first extracted feature and the third extracted feature to obtain a first calculation result includes:

[0226] (1.1) inputting the first extracted feature into the output layer, and inputting the third extracted feature into the output layer;

[0227] (1.2) The similarity between the first extracted feature and the third extracted feature is calculated at the output layer according to the noise contrast estimation loss function to obtain a first calculation result.

[0228] The first extracted feature may be considered as a feature related to the object finally extracted by the object tower, and the third extracted feature may be considered as a feature related to the pushed information.

[0229] The noise contrast estimation loss function can be used to calculate the similarity between the first extracted feature and the third extracted feature at the output layer to obtain a first calculation result. The noise estimation contrast loss function can be used to convert a multi-classification problem into a binary classification problem. For example, an object is related to multiple push information, which can be converted into whether the object is compatible with the currently input push information. This reduces the complexity of the comparison recommendation model in the output layer calculation and improves the computing performance of the object recommendation model.

[0230] The noise contrast estimation loss function is as follows:

[0231]

[0232] Among them, σ is the sigmoid function, Ti is the total sample set, Si is the negative sample set, Xi is the extracted features output by the object tower, and yi is the extracted features output by the item tower.

[0233] The first extracted feature and the third extracted feature may be input into the noise contrast estimation loss function to calculate a first calculation result.

[0234] In some implementations, the comparative recommendation model may also use a cosine loss function, and calculate the cosine distance value between the first extracted feature and the third extracted feature by using the cosine loss function as the first calculation result.

[0235] 304. Perform similarity calculation on the second extracted feature and the third extracted feature to obtain a second calculation result.

[0236] The second extracted feature may be considered as a feature related to the object finally extracted by the object tower, and the third extracted feature may be considered as a feature related to the pushed information.

[0237] A noise contrast estimation loss function may be used to calculate the similarity between the second extracted feature and the third extracted feature at the output layer to obtain a second calculation result.

[0238] In some implementations, the comparative recommendation model may also use a cosine loss function, and calculate the cosine distance value between the second extracted feature and the third extracted feature by using the cosine loss function as the second calculation result.

[0239] 305. Train the basic object recommendation model according to the first calculation result and the second calculation result to obtain a first object recommendation model.

[0240] After obtaining the first calculation result, the first calculation result can be compared with the preset loss function condition. For example, the preset loss function condition is a first preset loss range and a second preset loss range. The maximum value of the first preset loss range is less than the minimum value of the second preset range. When the first calculation result is in the first preset loss range, the first extracted feature and the third extracted feature are similar.

[0241] When the first calculation result is within the second preset loss range, the first extracted feature and the third extracted feature are not similar. Since the first training data is a positive sample, if the positive sample is related to the historical delivery information, and the result output by the trained object recommendation model is that the first extracted feature and the third extracted feature are not similar, it means that the object recommendation model needs to be further trained, such as by adjusting the model parameters of the object recommendation model, so that the finally trained object recommendation model can accurately predict the target object adapted to the push information in the input object group.

[0242] After obtaining the second calculation result, the second calculation result can be compared with the preset loss function condition. When the second calculation result is within the first preset loss range, the second extracted feature and the third extracted feature are not similar. Since the second training data is a negative sample, if the negative sample is not related to the historical delivery information, and the result output by the trained object recommendation model is that the second extracted feature and the third extracted feature are similar, it means that the object recommendation model needs to be further trained, such as by adjusting the model parameters of the object recommendation model, so that the finally trained object recommendation model can accurately determine whether the object and the push information match.

[0243] It can be seen from the above that after the basic object recommendation model is initially trained, the obtained first object model can preliminarily identify the target object to which the push information is adapted in the object group.

[0244] In step 140, the input object data corresponding to the object group to be screened and the target push information data of the target push information pushed by the promoter are input into the first object recommendation model, and a third object group corresponding to the target push information is screened out from the object group to be screened.

[0245] After determining the first object recommendation model, the input object data of the object group to be screened can be obtained. The object group to be screened can be understood as some object groups currently served by the service provider. The service provider can perform preliminary object screening of the target push information for the object group to be screened.

[0246] In some embodiments, the input object data of the group of objects to be screened includes the object data of each object in the group of objects to be screened, and the input push information data of the push information browsed by each object in the group of objects to be screened. The object data of each object in the group of objects to be screened includes age, the region, education level, and preferences, etc. The input push information data of each object in the group of objects to be screened includes a scheduling identifier, a push information identifier, and an industry category identifier, etc.

[0247] It is also possible to obtain the target push information data of the target push information pushed by the promoter, such as the target push information data including the schedule identifier, the push information identifier and the industry category identifier, etc. Among them, the target push information data can be understood as the push information currently released by the promoter.

[0248] It can be understood that, compared with the basic object recommendation model, only some setting parameters of the first object recommendation model have changed, but the structure of the model has not changed, that is, the first object recommendation model and the basic object recommendation model have the same structure.

[0249] Furthermore, the object data to be screened of each object in the screening object group is input into the first sub-object tower of the first object recommendation model, the input push information data of each object in the screening object group is input into at least one second sub-object tower of the first object recommendation model, and the target push information data is input into the item tower of the first object recommendation model, so as to use the first object recommendation model to determine a third object group that is compatible with the target push information in the object group to be screened.

[0250] The third object group may be understood as the target group that is initially screened out by the first object recommendation model and that meets the target push information requirements.

[0251] In step 150 , a fourth target group selected by the promoter and a fifth target group not selected are obtained from the third target group.

[0252] It is understandable that the third object group is merely a group preliminarily screened out by the service provider that is compatible with the target push information, and the third object group needs to be sent to the promoter for selection. For example, the service provider can send the third object group to the demand-side platform, and then the promoter can select the objects in the third object group on the demand-side platform.

[0253] In some implementations, after the third object group is acquired, the third object group may be sent to the promoter, and the promoter determines the selected fourth object group and the unselected fifth object group from the third object group.

[0254] The fourth target group selected by the promoter and the fifth target group not selected by the promoter are obtained from the third target group, including:

[0255] (1.1) Sending third-party target groups to the promoter;

[0256] (1.2) receiving a fourth target group selected by the promoter from the third target group;

[0257] (1.3) A fifth target group not selected by the receiving promoter from the third target group.

[0258] The promoter has set up corresponding screening criteria internally, and not all the objects in the third object group are considered by the promoter to be suitable for the target push information. The promoter can select a fourth object group from the third object group, and the fourth object group can be understood as the object group that the promoter considers to be suitable for the target push information, that is, the fourth object group is the audience of the target push information.

[0259] The group not selected by the promoter from the third target group is the fifth target group. The fifth target group can be considered as a target group that is not compatible with the target push information, that is, the fifth target group is not the audience of the target push information.

[0260] The service provider may receive the fourth target group selected by the promoter from the third target group according to the demand-side platform, and receive the fifth target group not selected by the promoter from the third target group.

[0261] In step 160, first incremental training data corresponding to the fourth object group and second incremental training data corresponding to the fifth object group are obtained.

[0262] Among them, in order to make the object recommendation model more suitable for the target push information of the promoter, the first object recommendation model can continue to be trained. The training process can be understood as an incremental training of the first object recommendation model, so that the trained first object recommendation model is more suitable for the target push information of the promoter.

[0263] In some implementations, first incremental training data corresponding to the fourth subject group and second incremental training data corresponding to the fifth subject group may be obtained. The method of obtaining the first incremental training data corresponding to the fourth subject group and the second incremental training data corresponding to the fifth subject group includes:

[0264] (1.1) Acquire the third object data of each object in the fourth object group, and the third push information data of the push information browsed by each object in the fourth object group;

[0265] (1.2) generating first incremental training data corresponding to the fourth object group according to the third object data and the third push information data of each object in the fourth object group;

[0266] (1.3) acquiring fourth object data of each object in the fifth object group, and fourth push information data of push information browsed by each object in the fifth object group;

[0267] (1.4) Generate second incremental training data corresponding to the sample object group based on the fourth object data and the fourth push information data of each object in the fifth object group.

[0268] The third object data of each object in the fourth object group includes: basic attributes of the object, such as age, region, education level and preference, etc. The third push information data of the push information browsed by each object in the fourth object group includes: a schedule identifier of the third push information, a push information identifier and an industry category identifier, etc.

[0269] Then, first incremental training data corresponding to the fourth object group is generated according to the third object data and the third push information data of each object in the fourth object group. The first incremental training data can be understood as positive sample data for continuing to train the first object recommendation model.

[0270] The fourth object data of each object in the fifth object group includes: basic attributes of the object, such as age, region, education level and preference, etc. The fourth push information data of the push information browsed by each object in the fifth object group includes: a schedule identifier of the fourth push information, a push information identifier and an industry category identifier, etc.

[0271] Then, according to the fourth object data and the fourth push information data of each object in the fifth object group, second incremental training data corresponding to the sample object group is generated. The second incremental training data can be understood as negative sample data for continuing to train the first object recommendation model.

[0272] It can be understood that, during the initial delivery of the target push information corresponding to the promoter by the service provider, only the second incremental training data corresponding to the fifth object group fed back by the promoter is used as negative sample data. Since the amount of the second incremental training data is too small, it is insufficient to support the training of the first object recommendation model.

[0273] In some embodiments, considering that during the initial delivery of the target push information, some objects in the delivered object group have a higher number of exposures to the target push information, but these objects have not clicked on or browsed the detailed content of the target push information, these object groups with higher exposure times can be used as negative sample populations.

[0274] Specifically, before training the first object recommendation model according to the first incremental training data, the second incremental training data and the target push information data to obtain the second object recommendation model, the method further includes:

[0275] (2.1) determining, among the target groups corresponding to the target push information and having been delivered, a sixth target group for which the number of exposures of the target push information is higher than a preset number of exposures;

[0276] (2.2) Obtain the third incremental training data corresponding to the sixth object group.

[0277] For example, the preset exposure times is 7. If the exposure times of some objects in the delivered object groups are higher than 7, such object groups are determined as the sixth object group. Then, the third incremental training data corresponding to the sixth object group is obtained.

[0278] Among them, the third incremental training data corresponding to the sixth object group includes the sixth object data corresponding to each object in the sixth object group, and the fifth push information data of the push information browsed by each object in the sixth object group. Among them, the sixth object data corresponding to each object in the sixth object group: the basic attributes of the object, such as age, the region, education level and preferences, etc. The fifth push information data of the push information browsed by each object in the sixth object group includes: the scheduling identifier of the fourth push information, the push information identifier and the industry category identifier, etc. The third incremental training data can be understood as negative sample data for training the first object recommendation model.

[0279] In step 170, the first object recommendation model is trained according to the first incremental training data, the second incremental training data and the target push information data to obtain a second object recommendation model.

[0280] When the amount of the second incremental training data is sufficient, the first object recommendation model can be trained according to the first incremental training data, the second incremental training data and the target push information data to obtain the second object recommendation model.

[0281] For example, the first incremental training data can be input into the object tower to output the fourth extracted feature, and the second incremental training data can be input into the object tower to output the fifth extracted feature. The fourth extracted feature and the seventh extracted feature are similarly calculated to obtain a third calculation result. The fifth extracted feature and the seventh extracted feature are similarly calculated to obtain a fourth calculation result. The first object recommendation model is trained according to the third calculation result and the fourth calculation result to obtain a second object recommendation model.

[0282] It should be noted that the training method for the first object recommendation model is similar to the training method for the basic object recommendation model.

[0283] In some implementations, the third incremental training data can also be used as negative sample data to train the first object recommendation model. Then, the first object recommendation model is trained according to the first incremental training data, the second incremental training data, the third incremental training data, and the target push information data to obtain a second object recommendation model.

[0284] See also Figure 7 , Figure 7 : is a flow chart of the training of the first object recommendation model provided in the embodiment of the present application. Further, the first object recommendation model is trained according to the first incremental training data, the second incremental training data, the third incremental training data and the target push information data to obtain the second object recommendation model, including:

[0285] In step 401, the first incremental training data is input into the object tower to output the fourth extracted feature, the second incremental training data is input into the object tower to output the fifth extracted feature, and the third incremental training data is input into the object tower to output the sixth extracted feature.

[0286] The structure of the first object recommendation model has not changed. The structure of the first object recommendation model is the same as that of the basic object recommendation model. Figure 5 And the above detailed description of the basic object recommendation model.

[0287] The first incremental data includes the first incremental object data corresponding to each object in the fourth object group, and also includes the first incremental push information data of the push information browsed by each object in the fourth object group. The first incremental object data includes basic attributes of the object, such as age, region, education level and preferences. The first incremental push information data includes a scheduling identifier, a push information identifier and an industry category identifier.

[0288] The first incremental training data is input into the first sub-object tower in the object tower, the first incremental push information data is input into at least one second sub-object tower, the results output by the first sub-object tower and at least one second sub-object tower are respectively input into the multi-layer perceptron, and the multi-layer perceptron outputs the fourth extracted feature corresponding to the first incremental training data.

[0289] The second incremental data includes the second incremental object data corresponding to each object in the fifth object group, and also includes the second incremental push information data of the push information browsed by each object in the fifth object group. The second incremental object data includes the basic attributes of the object, such as age, region, education level and preferences. The second incremental push information data includes a scheduling identifier, a push information identifier and an industry category identifier.

[0290] The second incremental training data is input into the first sub-object tower in the object tower, the second incremental push information data is input into at least one second sub-object tower, the results output by the first sub-object tower and at least one second sub-object tower are respectively input into the multi-layer perceptron, and the multi-layer perceptron outputs the fifth extracted feature corresponding to the second incremental training data.

[0291] The third incremental data includes the third incremental object data corresponding to each object in the sixth object group, and also includes the third incremental push information data of the push information browsed by each object in the sixth object group. The third incremental object data includes basic attributes of the object, such as age, region, education level and preferences. The third incremental push information data includes a scheduling identifier, a push information identifier and an industry category identifier.

[0292] The third incremental training data is input into the first sub-object tower in the object tower, the third incremental push information data is input into at least one second sub-object tower, the results output by the first sub-object tower and at least one second sub-object tower are respectively input into the multilayer perceptron, and the multilayer perceptron outputs the fifth extracted feature corresponding to the sixth incremental training data.

[0293] In step 402, the target push information data is input into the item tower, and the seventh extraction feature is output.

[0294] In some embodiments, after the target push information data is acquired, the target push information data is input into an item tower, the item tower includes an embedding layer, and the target push information data is feature extracted in the item tower through the embedding layer, thereby obtaining a seventh extracted feature.

[0295] Specifically, the data such as the schedule identifier, the push information identifier, and the industry category identifier in the target push information data can be input into the item tower to obtain the seventh extraction feature.

[0296] In step 403, similarity calculation is performed on the fourth extracted feature and the seventh extracted feature to obtain a third calculation result.

[0297] Specifically, the fourth extracted feature can be input into the output layer, and the seventh extracted feature can be input into the output layer. Then, the similarity between the fourth extracted feature and the seventh extracted feature is calculated in the output layer according to the noise contrast estimation loss function to obtain a third calculation result.

[0298] In step 404, similarity calculation is performed on the fifth extracted feature and the seventh extracted feature to obtain a fourth calculation result.

[0299] Specifically, the fifth extracted feature can be input into the output layer, and the seventh extracted feature can be input into the output layer. Then, the similarity between the fifth extracted feature and the seventh extracted feature is calculated in the output layer according to the noise contrast estimation loss function to obtain a fourth calculation result.

[0300] In step 405, similarity calculation is performed on the sixth extracted feature and the seventh extracted feature to obtain a fifth calculation result.

[0301] Specifically, the sixth extracted feature can be input into the output layer, and the seventh extracted feature can be input into the output layer. Then, the similarity between the sixth extracted feature and the seventh extracted feature is calculated in the output layer according to the noise contrast estimation loss function to obtain a fifth calculation result.

[0302] In step 406, the first object recommendation model is trained according to the third calculation result, the fourth calculation result and the fifth calculation result to obtain a second object recommendation model.

[0303] It is understandable that after obtaining the third calculation result, the fourth calculation result and the fifth calculation result, it can be determined whether the third calculation result, the fourth calculation result and the fifth calculation result meet the preset loss function conditions. If any of the third calculation result, the fourth calculation result and the fifth calculation result does not meet the preset loss function conditions, it means that the parameters of the first object recommendation model need to be further adjusted. Until the trained first object recommendation model can accurately identify the object that is compatible with the target push information, the second object recommendation model is obtained.

[0304] The second object recommendation model is more suitable for the target push information of the promoter. After using the second object recommendation model, the second object recommendation model can better screen the target objects that are adapted to the target push information from the object group served by the service provider. For example, these target objects are the audiences of the target push information. In this way, the proportion of the target objects in the object group served by the service provider can be increased. In other words, the trained second object recommendation model in the embodiment of the present application can automatically and accurately determine the target object group that the push information is adapted to from the object group.

[0305] As can be seen from the above, in an embodiment of the present application, by obtaining the intersection object group between the first object group and the second object group of the promoter, and obtaining a sample object group other than the intersection object group in the first object group; obtaining first training data corresponding to the intersection object group, second training data corresponding to the sample object group, and historical push information data of the historical push information that has been released; training the basic object recommendation model according to the first training data, the second training data, and the historical push information data to obtain a first object recommendation model; inputting the input object data corresponding to the object group to be screened and the target push information data of the target push information pushed by the promotion platform into the first object recommendation model, and screening out the third object group corresponding to the target push information from the object group to be screened; obtaining the fourth object group selected by the promoter and the fifth object group not selected in the third object group; obtaining the first incremental training data corresponding to the fourth object group and the second incremental training data corresponding to the fifth object group; training the first object recommendation model according to the first incremental training data, the second incremental training data, and the target push information data to obtain a second object recommendation model. In this way, through the interaction between the promoter and the service provider, the common intersection object group and the sample object group without intersection are determined, and the basic object recommendation is preliminarily trained in combination with the historical push information that has been delivered to obtain a first object recommendation model. The first object recommendation model is used to preliminarily predict the third object group corresponding to the target push information, and the third object group is divided into a selected fourth object group and an unselected fifth object group according to the needs of the promoter. The first object recommendation model is trained for the second time to obtain a trained second object recommendation model. Compared with the scheme of manually circling or selecting the target object group by single click rate, the trained second object recommendation model of the present application is more suitable for the promoter's target push information, and can automatically and accurately determine the target object group that is compatible with the target push information among the object groups.

[0306] Please continue reading Figure 8 , Figure 8 1 is another flow chart of the object recommendation model training method provided in an embodiment of the present application. The object recommendation model training method may include the following steps:

[0307] In step 501, an intersection object group between a first object group and a second object group of a promoter is obtained, and a sample object group other than the intersection object group is obtained from the first object group.

[0308] It is understandable that before training the basic object recommendation model, it is also necessary to obtain the training samples corresponding to the basic object recommendation model, where the training samples include positive samples and negative samples. Considering that the object recommendation model serves the promoter, the intersection object group between the local first object group and the promoter's second object group can be obtained. Specifically, the intersection object group can be obtained by intersecting the privacy sets, thereby ensuring that when the intersection object group is obtained, no user privacy will be leaked between the service provider and the promoter. Since the intersection object group is also the object group saved by the promoter, the data corresponding to the intersection object group can be understood as the positive sample in the training sample of the basic object recommendation model.

[0309] After obtaining the intersection object group, a sample object group other than the intersection object group can be obtained from the first object group. For example, in the group other than the intersection object group in the first object group, a portion of the sample object group can be screened out by random screening, and the data corresponding to the sample object group can be understood as a negative sample in the training sample of the object recommendation model.

[0310] In step 502, the first object data of each object in the intersection object group and the first push information data of the push information browsed by each object in the intersection object group are obtained.

[0311] After obtaining the intersection object group, data corresponding to each object in the intersection object group can be obtained, and then the data corresponding to each object in the intersection object group is used to generate first training data. After obtaining the sample object group, data corresponding to each object in the sample object group can be obtained, and then the data corresponding to each object in the sample object group is used to generate second training data.

[0312] In some implementations, first object data of each object in the intersection object group and first push information data of push information browsed by each object in the intersection object group may be obtained;

[0313] Generate first training data corresponding to the intersection object group according to the first object data and the first push information data of each object in the intersection object group;

[0314] Acquire second object data of each object in the sample object group, and second push information data of push information browsed by each object in the sample object group;

[0315] Second training data corresponding to the sample object group is generated according to the second object data and the second push information data of each object in the sample object group.

[0316] The first object data of each object in the intersection object group includes basic attributes of the object, such as age, region, education level, preference, etc. These basic attributes of the object are obtained after authorization by the object, or are basic attributes actively input by the object when using the service of the service provider.

[0317] The first push information data of the push information browsed by each object in the intersection object group may also include: a schedule identifier of the push information, a push information identifier, an industry category identifier, and the like.

[0318] After the first object data and the first push information data are acquired, the two may be determined as first training data. The first training data may be understood as positive sample data corresponding to the basic object recommendation model during training.

[0319] The second object data of each object in the sample object group includes basic attributes of the object, such as age, region, education level, preference, etc. These basic attributes of the object are obtained after authorization by the object, or are basic attributes actively input by the object when using the service of the service provider.

[0320] The second push information data of the push information browsed by each object in the sample object group may also include: a schedule identifier of the push information, a push information identifier, an industry category identifier, and the like.

[0321] After the second object data and the second push information data are acquired, the two can be determined as second training data. The second training data can be understood as negative sample data corresponding to the basic object recommendation model during training.

[0322] In some implementations, it is also necessary to obtain historical push information data of historical push information that has been delivered. The delivered historical push information may be push information within a preset historical time period before the current time, and the historical push information may also be push information that the promoter has delivered to the service provider. The historical push information data may include: a schedule identifier of the push information, a push information identifier, and an industry category identifier, etc.

[0323] In step 503, the first training data is input into the object tower, and the first extracted features are output, and the second training data is input into the object tower, and the second extracted features are output.

[0324] like Figure 5 As shown, the object tower includes multiple sub-object towers, specifically, the object tower includes a first sub-object tower and at least one second sub-object tower. Each sub-object tower is correspondingly provided with an embedding layer. When each sub-object tower inputs data, the embedding layer can extract the features of the data, thereby outputting vector features related to the input data.

[0325] The first object data in the first training data can be input into the first sub-object tower to output the first feature; the first push information data in the first training data can be input into at least one second sub-object tower to output the second feature; the first feature and the second feature can be input into the multi-layer perceptron to output the first extracted feature.

[0326] The second object data in the second training data can be input into the first sub-object tower to output the third feature; the second push information data in the second training data can be input into at least one second sub-object tower to output the fourth feature; the third feature and the fourth feature can be input into the multilayer perceptron to output the second extracted feature.

[0327] In step 504, the historical push information data is input into the item tower, and the third extracted feature is output.

[0328] In some embodiments, after the historical push information data is acquired, the historical push information data is input into an item tower, and the item tower includes an embedding layer, in which the historical push information data is feature extracted through the embedding layer to obtain a third extracted feature.

[0329] Specifically, data such as the schedule identifier, the push information identifier, and the industry category identifier in the historical push information data can be input into the item tower to obtain the third extracted feature.

[0330] In step 505, similarity calculation is performed on the first extracted feature and the third extracted feature to obtain a first calculation result.

[0331] The first extracted feature may be considered as a feature related to the object finally extracted by the object tower, and the third extracted feature may be considered as a feature related to the pushed information.

[0332] A noise contrast estimation loss function may be used to calculate the similarity between the first extracted feature and the third extracted feature at the output layer to obtain a first calculation result.

[0333] In step 506, similarity calculation is performed on the second extracted feature and the third extracted feature to obtain a second calculation result.

[0334] The second extracted feature can be considered as a feature related to the object finally extracted by the object tower, and the third extracted feature can be considered as a feature related to the pushed information.

[0335] A noise contrast estimation loss function may be used to calculate the similarity between the second extracted feature and the third extracted feature at the output layer to obtain a second calculation result.

[0336] In step 507, the basic object recommendation model is trained according to the first calculation result and the second calculation result to obtain a first object recommendation model.

[0337] After obtaining the first calculation result, the first calculation result can be compared with the preset loss function condition. For example, the preset loss function condition is a first preset loss range and a second preset loss range. The maximum value of the first preset loss range is less than the minimum value of the second preset range. When the first calculation result is in the first preset loss range, the first extracted feature and the third extracted feature are similar.

[0338] When the first calculation result is within the second preset loss range, the first extracted feature and the third extracted feature are not similar. Since the first training data is a positive sample, if the positive sample is related to the historical delivery information, and the result output by the trained object recommendation model is that the first extracted feature and the third extracted feature are not similar, it means that the object recommendation model needs to be further trained, such as by adjusting the model parameters of the object recommendation model, so that the finally trained object recommendation model can accurately predict the target object adapted to the push information in the input object group.

[0339] After obtaining the second calculation result, the second calculation result can be compared with the preset loss function condition. When the second calculation result is within the first preset loss range, the second extracted feature and the third extracted feature are not similar. Since the second training data is a negative sample, if the negative sample is not related to the historical delivery information, and the result output by the trained object recommendation model is that the second extracted feature and the third extracted feature are similar, it means that the object recommendation model needs to be further trained, such as by adjusting the model parameters of the object recommendation model, so that the finally trained object recommendation model can accurately determine whether the object and the push information match.

[0340] It can be seen from the above that after the basic object recommendation model is initially trained, the obtained first object model can preliminarily identify the target object to which the push information is adapted in the object group.

[0341] In step 508, the input object data corresponding to the object group to be screened and the target push information data of the target push information pushed by the promotion platform are input into the first object recommendation model, and a third object group corresponding to the target push information is screened out from the object group to be screened.

[0342] In some embodiments, the input object data of the group of objects to be screened includes the object data of each object in the group of objects to be screened, and the input push information data of the push information browsed by each object in the group of objects to be screened. The object data of each object in the group of objects to be screened includes age, the region, education level, and preferences, etc. The input push information data of each object in the group of objects to be screened includes a scheduling identifier, a push information identifier, and an industry category identifier, etc.

[0343] It is also possible to obtain the target push information data of the target push information pushed by the promoter, such as the target push information data including the schedule identifier, the push information identifier and the industry category identifier, etc. Among them, the target push information data can be understood as the push information currently released by the promoter.

[0344] It can be understood that, compared with the basic object recommendation model, only some setting parameters of the first object recommendation model have changed, but the structure of the model has not changed, that is, the first object recommendation model and the basic object recommendation model have the same structure.

[0345] Furthermore, the object data to be screened of each object in the screening object group is input into the first sub-object tower of the first object recommendation model, the input push information data of each object in the screening object group is input into at least one second sub-object tower of the first object recommendation model, and the target push information data is input into the item tower of the first object recommendation model, so as to use the first object recommendation model to determine a third object group that is compatible with the target push information in the object group to be screened.

[0346] The third object group may be understood as the target group that is initially screened out by the first object recommendation model and that meets the target push information requirements.

[0347] In step 509, a fourth target group selected by the promoter and a fifth target group not selected are obtained from the third target group.

[0348] It is understandable that the third object group is merely a group preliminarily screened out by the service provider that is compatible with the target push information, and the third object group needs to be sent to the promoter for selection. For example, the service provider can send the third object group to the demand-side platform, and then the promoter can select the objects in the third object group on the demand-side platform.

[0349] The promoter has set up corresponding screening criteria internally, and not all the objects in the third object group are considered by the promoter to be suitable for the target push information. The promoter can select a fourth object group from the third object group, and the fourth object group can be understood as the object group that the promoter considers to be suitable for the target push information, that is, the fourth object group is the audience of the target push information.

[0350] The group not selected by the promoter from the third target group is the fifth target group. The fifth target group can be considered as a target group that is not compatible with the target push information, that is, the fifth target group is not the audience of the target push information.

[0351] In step 510, first incremental training data corresponding to the fourth object group and second incremental training data corresponding to the fifth object group are obtained.

[0352] In some embodiments, the third object data of each object in the fourth object group and the third push information data of the push information browsed by each object in the fourth object group can be obtained; the first incremental training data corresponding to the fourth object group can be generated according to the third object data and the third push information data of each object in the fourth object group; the fourth object data of each object in the fifth object group and the fourth push information data of the push information browsed by each object in the fifth object group can be obtained; and the second incremental training data corresponding to the sample object group can be generated according to the fourth object data and the fourth push information data of each object in the fifth object group.

[0353] In step 511, a sixth object group whose exposure times of the target push information are higher than a preset exposure times is determined among the delivered object groups corresponding to the target push information, and third incremental training data corresponding to the sixth object group is obtained.

[0354] In some embodiments, considering that during the initial delivery of the target push information, some objects in the delivered object group have a higher number of exposures to the target push information, but these objects have not clicked on or browsed the detailed content of the target push information, these object groups with higher exposure times can be used as negative sample populations.

[0355] A sixth object group whose exposure times of the target push information are higher than the preset exposure times can be determined in the delivered object group corresponding to the target push information; and the third incremental training data corresponding to the sixth object group is obtained. The third incremental training data corresponding to the sixth object group includes the sixth object data corresponding to each object in the sixth object group and the fifth push information data of the push information browsed by each object in the sixth object group.

[0356] In step 512, the first object recommendation model is trained according to the first incremental training data, the second incremental training data, the third incremental training data and the target push information data to obtain a second object recommendation model.

[0357] It can be understood that the structure of the first object recommendation model is the same as that of the basic object recommendation model and has not changed.

[0358] The first incremental training data can be input into the object tower to output the fourth extracted feature, the second incremental training data can be input into the object tower to output the fifth extracted feature, and the third incremental training data can be input into the object tower to output the sixth extracted feature; the target push information data can be input into the item tower to output the seventh extracted feature; the fourth extracted feature and the seventh extracted feature are similarly calculated to obtain a third calculation result; the fifth extracted feature and the seventh extracted feature are similarly calculated to obtain a fourth calculation result; the sixth extracted feature and the seventh extracted feature are similarly calculated to obtain a fifth calculation result; the first object recommendation model is trained according to the third calculation result, the fourth calculation result and the fifth calculation result to obtain a second object recommendation model.

[0359] From the above content, it can be seen that by training the basic object recommendation model for the first time, the first object recommendation model is obtained, and then the first object recommendation model is trained for the second time to obtain the second object recommendation model. The second object recommendation model is more suitable for target push information and can automatically and accurately determine the target object group that the target push information is adapted to in the object group.

[0360] Please continue reading Fig. 9 , Fig. 9 It is a schematic diagram of a scenario corresponding to the object recommendation model provided in an embodiment of the present application.

[0361] Among them, before the promoter releases the push information, the service provider can train the basic object recommendation model.

[0362] First, the intersection object group between the first object group and the second object group of the promoter is obtained, and the sample object group other than the intersection object group is obtained in the first object group. For example, the intersection object group between the first object group and the second object group is determined by intersecting the privacy sets. Then, the sample object group other than the intersection object group is obtained in the first object group.

[0363] Then, the first training data corresponding to the intersection object group, the second training data corresponding to the sample object group, and the historical push information data of the historical push information that has been delivered are obtained, and the basic object recommendation model is trained according to the first training data, the second training data, and the historical push information data to obtain the first object recommendation model.

[0364] It can be understood that the first object recommendation model at this time can preliminarily determine the target object that the target push information released by the promoter is suitable for.

[0365] At this time, the input object data corresponding to the object group to be screened and the target push information data of the target push information pushed by the promotion platform are input into the first object recommendation model, and the third object group corresponding to the target push information is screened out from the object group to be screened. The object group to be screened can be understood as the object group that the service provider is using its service.

[0366] The service provider can then send the third object group to the promoter, and then obtain the fourth object group selected by the promoter and the fifth object group not selected by the promoter from the third object group. For example, the promoter can receive some objects in the third object group through the demand-side platform and roll back some objects in the third object group.

[0367] The first incremental training data corresponding to the fourth object group and the second incremental training data corresponding to the fifth object group are obtained. The first object recommendation model is trained according to the first incremental training data, the second incremental training data and the target push information data to obtain a second object recommendation model.

[0368] The second object recommendation model can more accurately screen out target objects that match the target push information from the input object group, thereby increasing the proportion of target objects that match the target push information in the input object group.

[0369] It can be seen from the above that in the present application, the trained object recommendation model can automatically and accurately determine the target object group to which the target push information is adapted from the object group.

[0370] After the second object recommendation model is trained, the embodiment of the present application further provides an object prediction method, which can accurately predict the target object to which the target push information is adapted in the input object group.

[0371] Please continue reading Fig.10 , Fig.10 : is a flow chart of an object prediction method provided in an embodiment of the present application. The object prediction method may include the following steps:

[0372] In step 601, the input object group and target push information are obtained.

[0373] After the second object recommendation model is trained, the second object recommendation model can perform real-time prediction on the input object group, thereby predicting the first target object that matches the target push information that the promoter currently needs to deliver from the input object group.

[0374] In some implementations, the service provider may obtain a currently inputted input object group, where the input object group may be understood as a group currently using the service provided by the service provider.

[0375] The service provider can obtain the target push information that the promoter currently needs to place. For example, the service provider can obtain the target push information that the promoter currently needs to place through the demand-side platform.

[0376] In step 602, first target input object data and target push information data of target push information corresponding to the input object group are obtained.

[0377] In some implementations, the first target input object data includes data such as age, region, education level, and preferences of each object in the input object group. The first target input object data is data obtained after authorization by the objects in the input object group.

[0378] The target push information data of the target push information includes data such as the schedule identifier, the push information identifier and the industry category identifier.

[0379] In step 603, the first target input object data and the target push information data are input into the trained second object recommendation model to predict the first target object corresponding to the target push information.

[0380] It is understandable that after training the basic object recommendation model multiple times, the second object recommendation model obtained is more suitable for the push information of the current promoter. After inputting the first target input object data and the target push information data into the trained second object recommendation model, the second object recommendation model can automatically and accurately predict the first target object adapted to the target push information in the input object group. The first target object can be understood as the audience corresponding to the target push information.

[0381] Specifically, the first target input object data can be input into the object tower of the second object recommendation model, and the target push information data can be input into the item tower of the second object recommendation model. After a series of feature extraction, the final processing is performed in the output layer of the second object recommendation model to obtain the output result. Finally, the first target object that is adapted to the target push information in the input object group is determined according to the output result.

[0382] In some implementations, after obtaining the first target object corresponding to the target push information, the method further includes:

[0383] (1.1) Obtain the total number of objects that require target push information and the importance of the target push information;

[0384] (1.2) Determine the number of objects in the input object group;

[0385] (1.3) determining the ratio of the number of first target objects to the number of objects;

[0386] (1.4) Generate constraints based on the total number of objects, importance, number of objects, and occupancy ratio;

[0387] (1.5) Based on the allocation algorithm and constraints, the target push information is delivered to the input object group.

[0388] After the first target object is determined, these first target objects may also be the target objects corresponding to other push information. If other push information preempts the first target object, the target push information cannot be delivered to the first target object, resulting in a shortage of the target push information during delivery. The shortage can be understood as the number of target push information delivered by the service provider does not meet the delivery number required by the target push information.

[0389] In order to solve this technical problem, an allocation algorithm is adopted in the embodiment of the present application. When the target push information is delivered online, the allocation algorithm can reasonably deliver it to the first target object, thereby preventing the situation where the target push information is insufficient.

[0390] Please also read Fig.11 , Fig.11 It is a schematic diagram of a scenario for applying the allocation algorithm provided in an embodiment of the present application.

[0391] Among them, after the promoter places an order, the order contract determines the amount of push information that the promoter needs to deliver, and then allocates the objects to different push information in the input object group through the allocation algorithm, thereby completing the delivery of the push information.

[0392] For example, the object recommendation model provided in the embodiment of the present application is used to determine the target object corresponding to each push information, and then the corresponding push information is allocated to each target object according to the allocation algorithm, thereby completing the delivery of the push information and reducing the shortage of the delivered push information.

[0393] like Fig.11 As shown, the allocation algorithm can deliver push information 1 and push information 3 to object 1 based on its characteristics: (35 years old, male, region A). The allocation algorithm can deliver push information 1 and push information 3 to object 4 based on its characteristics: (40 years old, female, region B).

[0394] In the process involving the delivery of multiple push information, the push information can be delivered reasonably through the allocation algorithm, thereby ensuring that the delivery volume of the push information is not insufficient.

[0395] The formula of the allocation algorithm is as follows:

[0396]

[0397]

[0398]

[0399]

[0400]

[0401] After introducing the Lagrange multiplier into the formula of the above allocation algorithm, the following formula is obtained:

[0402]

[0403] Among them, s i represents the number of objects in the input group, dj represents the number of objects required for the target push information, and s j Represents the number of remaining objects that need to be allocated for the target push information, p j Represents the penalty item when the target push information delivery does not meet the conditions, u j Indicates the number of objects to which the target push information has not yet been delivered. Represents the importance of the target advertisement, T j The total number of objects that need to push information to the target, t ij The ratio of the number of objects corresponding to the target push information to the number of objects in the input object group. ij is the quantity to be solved, that is, the proportion of objects that need to be allocated to the target push information in the total number of objects in the input object group. ij Represents the ideal allocation ratio of the number of objects in the input object group to the target push information. α, β, λ, γ, ψ are the dual variables of several constraint terms. Γ(i) is the set of objects that meet the target push information requirements.

[0404] Formula (1) represents the allocation target, that is, the target push information shortage rate is minimized. Formula (2) is the first constraint item, which specifically means that the number of allocated objects plus the number of unallocated objects corresponding to the target push information is greater than or equal to the number of objects required by the target push information. Formula (3) is the second constraint item, which specifically means that the sum of the proportions of the target push information should be less than or equal to 1. Formula (4) is the third constraint item, which specifically means that the product of the number of objects, the occupancy ratio, and the amount to be solved must be greater than or equal to the total number of objects required by the target object. Formula (5) is the fourth constraint item, which specifically means that the amount to be solved and the number of objects that have not yet been delivered by the target push information cannot be negative.

[0405] That is to say, in the above allocation algorithm, by adding the constraint condition corresponding to formula (4), the delivery of the target push information is achieved while ensuring that the first target object of the target push information occupies a proportion of the number of objects in the input object group, thereby ensuring that the target push information has a minimum shortage rate, or ensuring that the delivery of the target push information is not shortage.

[0406] As can be seen from the above content, in the embodiment of the present application, by obtaining the input object group and the target push information, obtaining the first target input object data corresponding to the input object group and the target push information data of the target push information, the first target input object data and the target push information data are input into the above-mentioned trained second object recommendation model, and the first target object corresponding to the target push information is predicted. The second object recommendation model can automatically and accurately determine the first target object that the target push information is adapted to in the input object group. At the same time, the target push information can also be reasonably allocated and delivered through the allocation algorithm, thereby ensuring that the target push information is not lacking in quantity during the delivery process.

[0407] In some embodiments, if the number of input object groups is small, the first target objects that are compatible with the target push information screened out using the second object recommendation model may be insufficient in number. For example, the target push information of the promoter requires an audience of 100, but the number of first target objects is only 70.

[0408] Through research, it is found that the click rate feature can also be used to determine the target objects to which the target push information is adapted. Therefore, the embodiment of the present application also provides a third object recommendation model, which increases the number of target objects to which the final target push information is adapted by adding the click rate feature.

[0409] See also Fig.12 , Fig.12 It is a structural diagram of the third object recommendation model provided in an embodiment of the present application.

[0410] The third object recommendation model is a trained model and can be used directly. The third object recommendation model includes an object tower, a first item tower, a second item tower, a first output layer, a second output layer and a gating network.

[0411] The second target input object data in the input object population can be input into the object tower, and the second target input object data includes the click rate corresponding to each object in the input object population, for example, the click rate is the click rate of each object in the input object population on the historical push information in the historical time period. The second target input object data also includes data such as age, region, education level and preference.

[0412] After the second target input object data is input into the object tower, the embedding layer of the object tower performs feature extraction to obtain output feature 1.

[0413] The first item tower can be used to input target push information data of target push information, such as target push information data including category identification, scheduling identification, push information identification, etc. The embedding layer of the first item tower performs feature extraction to obtain output feature 2.

[0414] The second item tower can be used to input target push information data of target push information, for example, the target push information data includes category identification, scheduling identification, push information identification, etc. The embedding layer of the second item tower performs feature extraction to obtain output feature 3.

[0415] The target push information data can be simultaneously input into the gated network, and the gated network can share the vector parameters corresponding to the embedding layers in the first item tower and the second item tower, and use the shared vector parameters to process the output feature 2, thereby obtaining the output feature 4. It can also be used to process the output feature 3, thereby obtaining the output feature 5.

[0416] Finally, input feature 1 and output feature 4 are input to the first output layer to obtain the first result, which can be considered as the first screening object group that is screened out from the input object group and is adapted to the target push information. Input feature 1 and output feature 5 are input to the second output layer to obtain the second result, which can be considered as the second screening object group that is screened out from the input object group and is adapted to the target push information. Finally, the first screening object group and the second screening object group are summed to obtain the second target object group.

[0417] It is understandable that, since the click-through rate feature is added to the third object recommendation model, the number of second target objects in the second target object group is greater than the number of first target objects, thereby ensuring that the number of target objects pushed by the service provider to the promoter is sufficient, that is, meeting the number of audiences required for the target push information.

[0418] See also Fig.13 , Fig.13 This is a schematic diagram of the structure of the object recommendation model training device provided in the embodiment of the present application. The meanings of the terms are the same as those in the above-mentioned object recommendation model training method, and the specific implementation details can refer to the description in the method embodiment.

[0419] The object recommendation model training device 700 may include: a first acquisition module 710 , a second acquisition module 720 , a first training module 730 , a screening module 740 , a third acquisition module 750 , a fourth acquisition module 760 , and a second training module 770 .

[0420] A first acquisition module 710 is used to acquire an intersection object group between the first object group and the second object group of the promoter, and acquire a sample object group other than the intersection object group in the first object group;

[0421] The second acquisition module 720 is used to acquire the first training data corresponding to the intersection object group, the second training data corresponding to the sample object group, and the historical push information data of the delivered historical push information;

[0422] A first training module 730 is used to train the basic object recommendation model according to the first training data, the second training data and the historical push information data to obtain a first object recommendation model;

[0423] The screening module 740 is used to input the input object data corresponding to the object group to be screened and the target push information data of the target push information pushed by the promotion platform into the first object recommendation model, and screen out a third object group corresponding to the target push information from the object group to be screened;

[0424] A third acquisition module 750 is used to acquire a fourth target group selected by the promoter and a fifth target group not selected from the third target group;

[0425] A fourth acquisition module 760, configured to acquire first incremental training data corresponding to a fourth object group and second incremental training data corresponding to a fifth object group;

[0426] The second training module 770 is used to train the first object recommendation model according to the first incremental training data, the second incremental training data and the target push information data to obtain a second object recommendation model.

[0427] In some implementations, the second acquisition module 720 further includes a first sub-acquisition module and a second sub-acquisition module.

[0428] A first sub-acquisition module, used to acquire first object data of each object in the intersection object group, and first push information data of push information browsed by each object in the intersection object group;

[0429] Generate first training data corresponding to the intersection object group according to the first object data and the first push information data of each object in the intersection object group;

[0430] A second sub-acquisition module, used to acquire second object data of each object in the sample object group, and second push information data of push information browsed by each object in the sample object group;

[0431] Second training data corresponding to the sample object group is generated according to the second object data and the second push information data of each object in the sample object group.

[0432] In some implementations, the second training module 770 includes a first sub-feature acquisition module, a second sub-feature acquisition module, and a sub-training module.

[0433] A first sub-feature acquisition module, used to input the first training data into the object tower, output the first extracted feature, and input the second training data into the object tower, output the second extracted feature;

[0434] The second sub-feature acquisition module is used to input the historical push information data into the item tower and output the third extracted feature;

[0435] A sub-training module, used for calculating the similarity between the first extracted feature and the third extracted feature to obtain a first calculation result;

[0436] Calculate the similarity between the second extracted feature and the third extracted feature to obtain a second calculation result;

[0437] The basic object recommendation model is trained according to the first calculation result and the second calculation result to obtain a first object recommendation model.

[0438] In some embodiments, the first sub-feature acquisition module is used to input the first object data in the first training data into the first sub-object tower and output the first feature;

[0439] Inputting the first push information data in the first training data into at least one second sub-object tower, and outputting a second feature;

[0440] The first feature and the second feature are input into a multilayer perceptron, and a first extracted feature is output.

[0441] In some embodiments, the first sub-feature acquisition module is used to input the second object data in the second training data into the first sub-object tower and output the third feature;

[0442] Inputting the second push information data in the second training data into at least one second sub-object tower, and outputting a fourth feature;

[0443] The third feature and the fourth feature are input into the multilayer perceptron, and the second extracted feature is output.

[0444] In some embodiments, the sub-training module is used to input the first extracted feature into the output layer and input the third extracted feature into the output layer;

[0445] Calculate the similarity between the first extracted feature and the third extracted feature at the output layer according to the noise contrast estimation loss function to obtain a first calculation result;

[0446] Inputting the second extracted feature into the output layer, and inputting the third extracted feature into the output layer;

[0447] The similarity between the second extracted feature and the third extracted feature is calculated at the output layer according to the noise contrast estimation loss function to obtain a second calculation result.

[0448] In some embodiments, the second sub-feature acquisition module is used to determine, in the push information data corresponding to at least one second sub-object tower, selected push information data that is at least partially identical to the historical push information data before inputting the historical push information data into the object tower and outputting the third extracted feature;

[0449] The vector parameter represented by the selected push information data in the second sub-object tower is determined as the target vector parameter in the item tower, and the corresponding selected push information data in the item tower is set as the target vector parameter.

[0450] In some implementations, the first acquisition module 710 includes a first sub-encryption module, a second sub-encryption module, and a sub-decryption module.

[0451] A first sub-encryption module, used to determine a first encryption result corresponding to first group information of a first object group;

[0452] A second sub-encryption module, used to determine a second encryption result corresponding to the second group information of the second target group of the promoter;

[0453] A sub-decryption module, used for performing intersection acquisition processing on the first encryption result and the second encryption result to obtain an intersection result;

[0454] The intersection result is decrypted to obtain an intersection object group between the first object group and the second object group.

[0455] In some implementations, the first sub-encryption module is used to encrypt first group information of the first object group to obtain first encrypted information;

[0456] Sending the first encrypted information to the promoter;

[0457] The second encrypted information returned by the promoter after encrypting the first encrypted information is received, and the second encrypted information is determined as the first encryption result.

[0458] In some embodiments, the second sub-encryption module is used to obtain third encrypted information after the promoter encrypts the second group information of the second target group;

[0459] The third encrypted information is encrypted to obtain fourth encrypted information, and the fourth encrypted information is determined as the second encryption result.

[0460] In some implementations, the third acquisition module 750 is used to send the third target group to the promoter;

[0461] receiving a fourth target group selected by the promoter from the third target group;

[0462] A fifth target group not selected by the promoter from the third target group is received.

[0463] In some embodiments, the fourth acquisition module 760 is used to acquire third object data of each object in the fourth object group, and third push information data of push information browsed by each object in the fourth object group;

[0464] generating first incremental training data corresponding to the fourth object group according to the third object data and the third push information data of each object in the fourth object group;

[0465] Acquire fourth object data of each object in the fifth object group, and fourth push information data of push information browsed by each object in the fifth object group;

[0466] Second incremental training data corresponding to the sample object group is generated according to the fourth object data and the fourth push information data of each object in the fifth object group.

[0467] In some embodiments, the second training module 770 is configured to input the first incremental training data into the object tower and output a fourth extracted feature, input the second incremental training data into the object tower and output a fifth extracted feature, and input the third incremental training data into the object tower and output a sixth extracted feature;

[0468] Input the target push information data into the item tower and output the seventh extraction feature;

[0469] Calculate the similarity between the fourth extracted feature and the seventh extracted feature to obtain a third calculation result;

[0470] Calculate the similarity between the fifth extracted feature and the seventh extracted feature to obtain a fourth calculation result;

[0471] Calculate the similarity between the sixth extracted feature and the seventh extracted feature to obtain a fifth calculation result;

[0472] The first object recommendation model is trained according to the third calculation result, the fourth calculation result and the fifth calculation result to obtain a second object recommendation model.

[0473] As can be seen from the above, in an embodiment of the present application, by obtaining the intersection object group between the first object group and the second object group of the promoter, and obtaining a sample object group other than the intersection object group in the first object group; obtaining first training data corresponding to the intersection object group, second training data corresponding to the sample object group, and historical push information data of the historical push information that has been released; training the basic object recommendation model according to the first training data, the second training data, and the historical push information data to obtain a first object recommendation model; inputting the input object data corresponding to the object group to be screened and the target push information data of the target push information pushed by the promotion platform into the first object recommendation model, and screening out the third object group corresponding to the target push information from the object group to be screened; obtaining the fourth object group selected by the promoter and the fifth object group not selected in the third object group; obtaining the first incremental training data corresponding to the fourth object group and the second incremental training data corresponding to the fifth object group; training the first object recommendation model according to the first incremental training data, the second incremental training data, and the target push information data to obtain a second object recommendation model. In this way, through the interaction between the promoter and the service provider, the common intersection object group and the sample object group without intersection are determined, and the basic object recommendation is preliminarily trained in combination with the historical push information that has been delivered to obtain a first object recommendation model. The first object recommendation model is used to preliminarily predict the third object group corresponding to the target push information, and the third object group is divided into a selected fourth object group and an unselected fifth object group according to the needs of the promoter. The first object recommendation model is trained for the second time to obtain a trained second object recommendation model. Compared with the scheme of manually circling or selecting the target object group by single click rate, the trained second object recommendation model of the present application is more suitable for the promoter's target push information, and can automatically and accurately determine the target object group that is compatible with the target push information among the object groups.

[0474] The specific implementation of each of the above modules can be found in the previous embodiments and will not be described in detail here.

[0475] See also Fig.14 , Fig.14 The following is a schematic diagram of the structure of the object prediction device provided in the embodiment of the present application. The meanings of the terms are the same as those in the above-mentioned object prediction method, and the specific implementation details can refer to the description in the method embodiment.

[0476] The object prediction device 800 may include: an input acquisition module 810 , a data acquisition module 820 , and a prediction module 830 .

[0477] An input acquisition module 810 is used to acquire input object groups and target push information;

[0478] The data acquisition module 820 is used to acquire the first target input object data and the target push information data of the target push information corresponding to the input object group;

[0479] The prediction module 830 is used to input the first target input object data and the target push information data into the second object recommendation model trained by the object recommendation model training method provided in the embodiment of the present application, and predict the first target object corresponding to the target push information.

[0480] In some embodiments, the prediction module 830 is further used to obtain the total number of objects requiring target push information and the importance of the target push information;

[0481] determining the number of objects in the input object population;

[0482] Determine the ratio of the number of first target objects to the number of objects;

[0483] Generate constraints based on total number of objects, importance, number of objects, and occupancy ratio;

[0484] Based on the allocation algorithm and constraints, the target push information is delivered to the input object group.

[0485] As can be seen from the above content, in the embodiment of the present application, by obtaining the input object group and the target push information, obtaining the first target input object data corresponding to the input object group and the target push information data of the target push information, the first target input object data and the target push information data are input into the above-mentioned trained second object recommendation model, and the first target object corresponding to the target push information is predicted. The second object recommendation model can automatically and accurately determine the first target object that the target push information is adapted to in the input object group. At the same time, the target push information can also be reasonably allocated and delivered through the allocation algorithm, thereby ensuring that the target push information is not lacking in quantity during the delivery process.

[0486] The embodiment of the present application also provides a computer device, which can be a server. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited to this. The terminal and the server can be directly or indirectly connected via wired or wireless communication, and this application does not limit this.

[0487] like Fig.15As shown, it shows a schematic diagram of the structure of the server involved in the embodiment of the present application, specifically:

[0488] The computer device may include one or more processing core processors 901, one or more computer-readable storage media memories 902, a power supply 903, an input unit 904 and other components. Those skilled in the art will appreciate that Fig.14 The computer device structure shown in the figure does not constitute a limitation on the computer device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently. Among them:

[0489] The processor 901 is the control center of the computer device. It uses various interfaces and lines to connect various parts of the entire computer device. By running or executing software programs and / or modules stored in the memory 902 and calling data stored in the memory 902, it executes various functions of the computer device and processes data, thereby monitoring the computer device as a whole. Optionally, the processor 901 may include one or more processing cores; optionally, the processor 901 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, object interface and application programs, etc., and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 901.

[0490] The memory 902 can be used to store software programs and modules. The processor 901 executes various functional applications and data processing by running the software programs and modules stored in the memory 902. The memory 902 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the server, etc. In addition, the memory 902 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices. Accordingly, the memory 902 may also include a memory controller to provide the processor 901 with access to the memory 902.

[0491] The computer device also includes a power supply 903 for supplying power to various components. Optionally, the power supply 903 can be logically connected to the processor 901 through a power management system, so as to manage charging, discharging, and power consumption through the power management system. The power supply 903 can also include any components such as one or more DC or AC power supplies, recharging systems, power failure detection circuits, power converters or inverters, and power status indicators.

[0492] The computer device may further include an input unit 904, which may be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to object settings and function controls.

[0493] Although not shown, the computer device may further include a display unit, etc., which will not be described in detail herein. Specifically in this embodiment, the processor 901 in the computer device will load the executable files corresponding to the processes of one or more application programs into the memory 902 according to the following instructions, and the processor 901 will run the application programs stored in the memory 902, thereby implementing the various method steps provided in the aforementioned embodiments, as follows:

[0494] Obtaining an intersection object group between the first object group and the second object group of the promoter, and obtaining a sample object group other than the intersection object group in the first object group;

[0495] Acquire the first training data corresponding to the intersection object group, the second training data corresponding to the sample object group, and the historical push information data of the delivered historical push information;

[0496] The basic object recommendation model is trained according to the first training data, the second training data and the historical push information data to obtain a first object recommendation model;

[0497] Inputting the input object data corresponding to the to-be-screened object group and the target push information data of the target push information pushed by the promotion platform into the first object recommendation model, and screening out a third object group corresponding to the target push information from the to-be-screened object group;

[0498] Acquire a fourth target group selected by the promoter and a fifth target group not selected from the third target group;

[0499] Acquire first incremental training data corresponding to the fourth object group and second incremental training data corresponding to the fifth object group;

[0500] The first object recommendation model is trained according to the first incremental training data, the second incremental training data and the target push information data to obtain a second object recommendation model.

[0501] In the above embodiments, the description of each embodiment has its own focus. For the part that is not described in detail in a certain embodiment, please refer to the detailed description of the object recommendation model training method above, which will not be repeated here.

[0502] As can be seen from the above, in an embodiment of the present application, by obtaining the intersection object group between the first object group and the second object group of the promoter, and obtaining a sample object group other than the intersection object group in the first object group; obtaining first training data corresponding to the intersection object group, second training data corresponding to the sample object group, and historical push information data of the historical push information that has been released; training the basic object recommendation model according to the first training data, the second training data, and the historical push information data to obtain a first object recommendation model; inputting the input object data corresponding to the object group to be screened and the target push information data of the target push information pushed by the promotion platform into the first object recommendation model, and screening out the third object group corresponding to the target push information from the object group to be screened; obtaining the fourth object group selected by the promoter and the fifth object group not selected in the third object group; obtaining the first incremental training data corresponding to the fourth object group and the second incremental training data corresponding to the fifth object group; training the first object recommendation model according to the first incremental training data, the second incremental training data, and the target push information data to obtain a second object recommendation model. In this way, through the interaction between the promoter and the service provider, the common intersection object group and the sample object group without intersection are determined, and the basic object recommendation is preliminarily trained in combination with the historical push information that has been delivered to obtain a first object recommendation model. The first object recommendation model is used to preliminarily predict the third object group corresponding to the target push information, and the third object group is divided into a selected fourth object group and an unselected fifth object group according to the needs of the promoter. The first object recommendation model is trained for the second time to obtain a trained second object recommendation model. Compared with the scheme of manually circling or selecting the target object group by single click rate, the trained second object recommendation model of the present application is more suitable for the promoter's target push information, and can automatically and accurately determine the target object group that is compatible with the target push information among the object groups.

[0503] A person of ordinary skill in the art will appreciate that all or part of the steps in the various methods of the above embodiments may be completed by instructions, or by controlling related hardware through instructions. The instructions may be stored in a computer-readable storage medium and loaded and executed by a processor.

[0504] To this end, an embodiment of the present application provides a computer-readable storage medium, in which a plurality of instructions are stored, and the instructions can be loaded by a processor to execute the steps in any object recommendation O-type training method provided in the embodiment of the present application. For example, the instructions can execute the following steps:

[0505] Obtaining an intersection object group between the first object group and the second object group of the promoter, and obtaining a sample object group other than the intersection object group in the first object group;

[0506] Acquire the first training data corresponding to the intersection object group, the second training data corresponding to the sample object group, and the historical push information data of the delivered historical push information;

[0507] The basic object recommendation model is trained according to the first training data, the second training data and the historical push information data to obtain a first object recommendation model;

[0508] Inputting the input object data corresponding to the to-be-screened object group and the target push information data of the target push information pushed by the promotion platform into the first object recommendation model, and screening out a third object group corresponding to the target push information from the to-be-screened object group;

[0509] Acquire a fourth target group selected by the promoter and a fifth target group not selected from the third target group;

[0510] Acquire first incremental training data corresponding to the fourth object group and second incremental training data corresponding to the fifth object group;

[0511] The first object recommendation model is trained according to the first incremental training data, the second incremental training data and the target push information data to obtain a second object recommendation model.

[0512] According to one aspect of the present application, a computer program product or a computer program is provided, the computer program product or the computer program includes computer instructions, the computer instructions 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 provided in various optional implementations provided in the above embodiments.

[0513] The specific implementation of the above operations can be found in the previous embodiments, which will not be described in detail here.

[0514] The computer-readable storage medium may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0515] Since the instructions stored in the computer-readable storage medium can execute the steps in any one of the network communication methods provided in the embodiments of the present application, the beneficial effects that can be achieved by any one of the network communication methods provided in the embodiments of the present application can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0516] The above is a detailed introduction to an object recommendation model training method and related methods, devices, media and equipment provided in an embodiment of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A method for training an object recommendation model, characterized in that: include: Acquire an intersection object group between the first object group and the second object group of the promoter, and acquire a sample object group other than the intersection object group in the first object group; Acquire the first training data corresponding to the intersection object group, the second training data corresponding to the sample object group, and the historical push information data of the delivered historical push information; Training a basic object recommendation model according to the first training data, the second training data, and the historical push information data to obtain a first object recommendation model; Inputting the input object data corresponding to the object group to be screened and the target push information data of the target push information pushed by the promoter into the first object recommendation model, and screening out a third object group corresponding to the target push information from the object group to be screened; Acquire a fourth target group selected by the promoter and a fifth target group not selected by the promoter from the third target group; Acquire first incremental training data corresponding to the fourth object group and second incremental training data corresponding to the fifth object group; The first object recommendation model is trained according to the first incremental training data, the second incremental training data, and the target push information data to obtain a second object recommendation model.

2. The object recommendation model training method according to claim 1, characterized in that: The obtaining of first training data corresponding to the intersection object group and second training data corresponding to the sample object group includes: Acquire first object data of each object in the intersection object group, and first push information data of push information browsed by each object in the intersection object group; generating first training data corresponding to the intersection object group according to the first object data and the first push information data of each object in the intersection object group; Acquire second object data of each object in the sample object group, and second push information data of push information browsed by each object in the sample object group; Second training data corresponding to the sample object group is generated according to the second object data and the second push information data of each object in the sample object group.

3. The object recommendation model training method according to claim 2, characterized in that: The basic object recommendation model includes an object tower and an item tower, and the basic object recommendation model is trained according to the first training data, the second training data and the historical push information data to obtain a first object recommendation model, including: Inputting the first training data into the object tower and outputting a first extracted feature, and inputting the second training data into the object tower and outputting a second extracted feature; Input the historical push information data into the item tower, and output the third extracted feature; Calculate the similarity between the first extracted feature and the third extracted feature to obtain a first calculation result; Calculate the similarity between the second extracted feature and the third extracted feature to obtain a second calculation result; The basic object recommendation model is trained according to the first calculation result and the second calculation result to obtain a first object recommendation model.

4. The object recommendation model training method according to claim 3, characterized in that: The basic object recommendation model further includes a multilayer perceptron, the output of the object tower is connected to the input of the multilayer perceptron, the object tower includes a first sub-object tower and at least one second sub-object tower, and the inputting the first training data into the object tower and outputting the first extracted feature includes: Inputting first object data in the first training data into the first sub-object tower, and outputting a first feature; Inputting the first push information data in the first training data into the at least one second sub-object tower, and outputting a second feature; Inputting the first feature and the second feature into the multilayer perceptron, and outputting a first extracted feature; The step of inputting the second training data into the object tower and outputting the second extracted features comprises: Inputting the second object data in the second training data into the first sub-object tower, and outputting a third feature; Inputting the second push information data in the second training data into the at least one second sub-object tower, and outputting a fourth feature; The third feature and the fourth feature are input into the multilayer perceptron, and a second extracted feature is output.

5. The object recommendation model training method according to claim 4, characterized in that: The basic object recommendation model further includes an output layer, the output of the multilayer perceptron is connected to the output layer, the output of the item tower is connected to the output layer, and the similarity calculation of the first extracted feature and the third extracted feature to obtain a first calculation result includes: Inputting the first extracted features into the output layer, and inputting the third extracted features into the output layer; Calculating the similarity between the first extracted feature and the third extracted feature at the output layer according to the noise contrast estimation loss function to obtain a first calculation result; The performing similarity calculation on the second extracted feature and the third extracted feature to obtain a second calculation result includes: Inputting the second extracted features into the output layer, and inputting the third extracted features into the output layer; The similarity between the second extracted feature and the third extracted feature is calculated at the output layer according to the noise contrast estimation loss function to obtain a second calculation result.

6. The object recommendation model training method according to claim 4, characterized in that: Before inputting the historical push information data into the item tower and outputting the third extracted feature, the method further includes: Determining, among the push information data corresponding to the at least one second sub-object tower, selected push information data that is at least partially identical to the historical push information data; The vector parameter represented by the selected push information data in the second sub-object tower is determined as the target vector parameter in the item tower, and the corresponding selected push information data in the item tower is set as the target vector parameter.

7. The object recommendation model training method according to claim 1, characterized in that: The step of obtaining an intersection object group between the first object group and the second object group of the promoter includes: Determine a first encryption result corresponding to first group information of a first object group; Determine a second encryption result corresponding to the second group information of the second target group of the promoter; Performing intersection acquisition processing on the first encryption result and the second encryption result to obtain an intersection result; The intersection result is decrypted to obtain an intersection object group between the first object group and the second object group.

8. The object recommendation model training method according to claim 7, characterized in that: The determining of the first encryption result corresponding to the first group information of the first object group includes: Encrypting first group information of a first object group to obtain first encrypted information; Sending the first encrypted information to the promoter; The second encrypted information returned by the promoter after encrypting the first encrypted information is received, and the second encrypted information is determined as the first encryption result.

9. The object recommendation model training method according to claim 7, characterized in that: The second encryption result corresponding to the second group information of the second target group of the promoter is determined, including: Acquire third encrypted information obtained by encrypting the second group information of the second target group by the promoter; The third encrypted information is encrypted to obtain fourth encrypted information, and the fourth encrypted information is determined as the second encryption result.

10. The object recommendation model training method according to claim 1, characterized in that: The step of obtaining the fourth target group selected by the promoter and the fifth target group not selected by the promoter from the third target group includes: Sending the third target group to the promoter; receiving a fourth target group selected by the promoter from the third target group; A fifth target group not selected by the promoter from the third target group is received.

11. The object recommendation model training method according to claim 1, characterized in that: The acquiring the first incremental training data corresponding to the fourth object group and the second incremental training data corresponding to the fifth object group includes: Acquire the third object data of each object in the fourth object group, and the third push information data of the push information browsed by each object in the fourth object group; generating first incremental training data corresponding to the fourth object group according to the third object data and the third push information data of each object in the fourth object group; Acquire fourth object data of each object in the fifth object group, and fourth push information data of push information browsed by each object in the fifth object group; Second incremental training data corresponding to the sample object group is generated according to the fourth object data and the fourth push information data of each object in the fifth object group.

12. The object recommendation model training method according to claim 1, characterized in that: Before training the first object recommendation model according to the first incremental training data, the second incremental training data, and the target push information data to obtain the second object recommendation model, the method further includes: Determining, among the delivered object groups corresponding to the target push information, a sixth object group for which the exposure number of the target push information is higher than the preset exposure number; Acquire third incremental training data corresponding to the sixth object group; The training of the first object recommendation model according to the first incremental training data, the second incremental training data and the target push information data to obtain a second object recommendation model includes: The first object recommendation model is trained according to the first incremental training data, the second incremental training data, the third incremental training data and the target push information data to obtain a second object recommendation model.

13. The object recommendation model training method according to claim 12, characterized in that: The first object recommendation model includes an object tower and an item tower, and the first object recommendation model is trained according to the first incremental training data, the second incremental training data, the third incremental training data, and the target push information data to obtain a second object recommendation model, including: Input the first incremental training data into the object tower, output a fourth extracted feature, input the second incremental training data into the object tower, output a fifth extracted feature, and input the third incremental training data into the object tower, output a sixth extracted feature; Input the target push information data into the item tower, and output the seventh extraction feature; Calculate the similarity between the fourth extracted feature and the seventh extracted feature to obtain a third calculation result; Calculate the similarity between the fifth extracted feature and the seventh extracted feature to obtain a fourth calculation result; Calculate the similarity between the sixth extracted feature and the seventh extracted feature to obtain a fifth calculation result; The first object recommendation model is trained according to the third calculation result, the fourth calculation result and the fifth calculation result to obtain a second object recommendation model.

14. An object prediction method, characterized in that: include: Get the input object group and target push information; Acquire first target input object data corresponding to the input object group and target push information data of the target push information; The first target input object data and the target push information data are input into a second object recommendation model trained according to the object recommendation model training method according to any one of claims 1 to 13, and a first target object corresponding to the target push information is predicted.

15. The object prediction method according to claim 14, characterized in that: After obtaining the first target object corresponding to the target push information, the method further includes: Obtain the total number of objects requiring the target push information and the importance of the target push information; determining the number of objects in the input object population; Determine the ratio of the number of the first target objects to the number of the objects; Generate a constraint condition according to the total number of objects, the importance, the number of objects, and the occupancy ratio; The target push information is delivered to the input object group based on the allocation algorithm and the constraint condition.

16. An object recommendation model training device, characterized in that: include: A first acquisition module, configured to acquire an intersection object group between a first object group and a second object group of the promoter, and to acquire a sample object group other than the intersection object group in the first object group; A second acquisition module is used to acquire the first training data corresponding to the intersection object group, the second training data corresponding to the sample object group, and the historical push information data of the delivered historical push information; A first training module, configured to train a basic object recommendation model according to the first training data, the second training data and the historical push information data to obtain a first object recommendation model; A screening module, configured to input the input object data corresponding to the to-be-screened object group and the target push information data of the target push information pushed by the promoter into the first object recommendation model, and screen out a third object group corresponding to the target push information from the to-be-screened object group; A third acquisition module, used for acquiring a fourth target group selected by the promoter and a fifth target group not selected by the promoter from the third target group; A fourth acquisition module, used to acquire the first incremental training data corresponding to the fourth object group and the second incremental training data corresponding to the fifth object group; The second training module is used to train the first object recommendation model according to the first incremental training data, the second incremental training data and the target push information data to obtain a second object recommendation model.

17. An object prediction device, characterized in that: include: An input acquisition module is used to obtain input object groups and target push information; A data acquisition module, used to acquire first target input object data corresponding to the input object group and target push information data of the target push information; A prediction module is used to input the first target input object data and the target push information data into a second object recommendation model trained according to any object recommendation model training method according to claims 1-13, and predict the first target object corresponding to the target push information.

18. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, which are suitable for being loaded by a processor to execute the steps in the object recommendation model training method described in any one of claims 1 to 13 or the steps in the object prediction method described in claim 14 or 15.

19. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the processor implements the steps of the object recommendation model training method described in any one of claims 1 to 13 or the steps of the object prediction method described in claim 14 or 15.

20. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the steps of the object recommendation model training method described in any one of claims 1 to 13 are implemented, or the steps of the object prediction method described in claim 14 or 15 are implemented.

Images