A data processing method, a computer device and a storage medium

Abstract

Embodiments of the present application disclose a data processing method, a computer device and a storage medium. The method comprises: obtaining an object set, the object set comprising N objects and object features of each of the N objects; determining a target object from the object set and obtaining an interactive object of the target object from the object set, the interactive object comprising the target object; obtaining an interactive closeness between the target object and the interactive object and advertisement interactive behavior information between the target object and candidate advertisement data; generating an association weight value between the target object and the interactive object according to the interactive closeness, the advertisement interactive behavior information and the object features of the interactive object; and generating an object feature vector of the target object according to the association weight value and the object features of the interactive object, the object feature vector of the target object being used to determine advertisement data for delivery to the target object. The present application can improve the delivery efficiency of the advertisement data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, and particularly relates to a data processing method, a computer device and a storage medium. BACKGROUND

[0002] With the continuous rise of various application platforms, the number of users in the application platforms is also increasing, and therefore, the operation of putting advertisement data in various application platforms to carry out advertisement publicity is more and more common.

[0003] In the prior art, the advertisement data is usually put by a designated advertisement putting user, and when the system puts the advertisement data of the advertisement putting user, the system usually needs to train a model through the related user data of the user who has interactive behaviors with the advertisement data historically put by the advertisement putting user, and subsequently, the model obtained by training can be used to determine which part of users to put the advertisement data of the advertisement putting user. Since there are usually multiple advertisement putting users, the system needs to train the model corresponding to each advertisement putting user respectively, and then the put crowd of the advertisement data of each advertisement putting user can be determined through the model corresponding to each advertisement putting user respectively obtained by training.

[0004] Therefore, in the prior art, a large amount of time is consumed to train the model corresponding to each advertisement putting user respectively, and then the put crowd of the advertisement data of each advertisement putting user can be determined through the model corresponding to each advertisement putting user respectively obtained by training, which leads to low efficiency of putting advertisement data. SUMMARY

[0005] In view of the above problems, the embodiments of the present application provide a data processing method, a computer device and a storage medium, which can improve the efficiency of putting advertisement data.

[0006] The embodiments of the present application provide a data processing method, which comprises the following steps.

[0007] An object set is obtained, the object set comprising N objects and object features of each object in the N objects, wherein N is a positive integer;

[0008] A target object is determined from the object set, and an interactive object of the target object is obtained from the object set; the interactive object comprises the target object;

[0009] An interactive closeness between the target object and the interactive object is obtained, and advertisement interactive behavior information between the target object and candidate advertisement data is obtained;

[0010] An association weight value between the target object and the interactive object is generated according to the interactive closeness, the advertisement interactive behavior information and the object features of the interactive object;

[0011] According to the association weight value and the object feature of the interactive object, an object feature vector of the target object is generated; the object feature vector of the target object is used to determine the advertisement data to be pushed to the target object.

[0012] The embodiment of the application provides a data processing device, comprising:

[0013] An acquisition module is configured to acquire an object set; the object set comprises N objects and object features of each object in the N objects, wherein N is a positive integer;

[0014] A determination module is configured to determine a target object from the object set and acquire interactive objects of the target object from the object set; the interactive objects comprise the target object;

[0015] The acquisition module is further configured to acquire an interactive closeness between the target object and the interactive objects and advertisement interactive behavior information between the target object and candidate advertisement data;

[0016] A generation module is configured to generate an association weight value between the target object and the interactive objects according to the interactive closeness, the advertisement interactive behavior information and the object features of the interactive objects;

[0017] The generation module is further configured to generate an object feature vector of the target object according to the association weight value and the object features of the interactive objects; the object feature vector of the target object is used to determine the advertisement data to be pushed to the target object.

[0018] The embodiment of the application provides a computer device, comprising a network interface, a processor and a memory; the network interface, the processor and the memory are connected; the network interface is configured to provide a data communication function; the memory is configured to store a computer program; and the processor is configured to call the computer program to execute part or all steps described in the embodiment of the application.

[0019] Correspondingly, the embodiment of the application provides a storage medium, the storage medium stores a computer program, the computer program comprises program instructions, and the program instructions are loaded and executed by one or more processors to execute the data processing method in the embodiment of the application.

[0020] Correspondingly, the application provides a computer program product or a computer program, the computer program product or the computer program comprises computer instructions, and the computer instructions are stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the data processing method in the aspect.

[0021] It can be seen that, in the embodiments of the present application, the association weight value between the target object and the interactive object is generated through the object features of the object, the interactive intimacy between the target object and the interactive object, and the advertising interactive behavior information between the target object and the candidate advertising data, and then the object feature vector of the target object is generated by using the association weight value and the interactive object features, wherein the interactive intimacy distinguishes different interactive objects of the target object, and the association weight value also reflects the importance of the object feature vectors of different interactive objects to the object feature vector of the target object. The introduction of advertising interactive behavior information to assist in generating the object feature vector can make the expression information of the object feature vector more rich and accurate, thereby effectively improving the accuracy of advertising data delivery. Since the advertising data to be delivered to the object can be directly determined according to the generated object feature vector, the delivery efficiency of the advertising data can also be improved. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0023] Figure 1 is an architecture schematic diagram of a data processing system provided by an embodiment of the present application;

[0024] Figure 2 is an interface schematic diagram of a crowd expansion extraction module provided by an embodiment of the present application;

[0025] Figure 3 is a flow schematic diagram of a data processing method provided by an embodiment of the present application;

[0026] Figure 4 is a structure schematic diagram of a language representation model provided by an embodiment of the present application;

[0027] Figure 5 is a structure schematic diagram of an attention mechanism provided by an embodiment of the present application;

[0028] Figure 6 is a scene schematic diagram of generating an object feature vector provided by an embodiment of the present application;

[0029] Figure 7 is a flow schematic diagram of a social diffusion algorithm provided by an embodiment of the present application;

[0030] Figure 8 is a flow schematic diagram of a data processing method provided by an embodiment of the present application;

[0031] Figure 9 is a flow diagram of a data processing method provided by an embodiment of the present application;

[0032] Figure 10 is a structural diagram of a data processing device provided by an embodiment of the present application;

[0033] Figure 11 is a structural diagram of a computer device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0034] In order to make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Therefore, the detailed description of the embodiments of the present application provided in the following drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application

[0035] In addition, the term "and / or" in this document is only to describe the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can represent the following three cases: A exists alone, A and B exist together, and B exists alone. In addition, the character " / " in this document generally represents an "or" relationship between the associated objects without special description.

[0036] The technical solutions provided by the present application belong to machine learning in artificial intelligence technology, and also involve big data in cloud infrastructure technology. The technologies involved in the data processing scheme provided by the present application will be briefly described below.

[0037] Cloud technology refers to a kind of hosting technology that unifies a series of resources such as hardware, software, network, etc. in a wide area network or local area network to realize data calculation, storage, processing and sharing. Cloud technology is a general term for network technology, information technology, integration technology, management platform technology, application technology, etc. based on cloud computing business model application, which can form a resource pool and be used on demand, flexibly and conveniently. Cloud computing technology will become an important support. The background service of the technical network system needs a large amount of computing and storage resources, such as video websites, picture websites and more portals. With the high development and application of the Internet industry, every item in the future may have its own identification mark and needs to be transmitted to the background system for logical processing. Different levels of data will be processed separately, and various industry data will need strong system support, which can only be realized through cloud computing.

[0038] Among them, the cloud infrastructure technology includes Big data, which refers to a collection of data that cannot be captured, managed and processed within a certain time range by conventional software tools, and is a massive, high-growth and diversified information asset that needs new processing mode to have stronger decision-making, insight discovery and process optimization capabilities. With the advent of the cloud era, Big data has attracted more and more attention, and Big data needs special technology to effectively process large amounts of data over time. The technologies suitable for Big data include large-scale parallel processing database, data mining, distributed file system, distributed database, cloud computing platform, Internet and scalable storage system. The object features and object feature vectors in the present application can be stored in the distributed file system in the cloud technology.

[0039] In addition, Artificial Intelligence (AI) is the use of digital computers or digital computer controlled machines to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which tries to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that machines have the functions of perception, reasoning and decision-making.

[0040] Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, both hardware and software level technology. Artificial intelligence basic technology generally includes, such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology and machine learning / deep learning and other fields.

[0041] Among them, machine learning (Machine Learning, ML) is a multi-disciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It is a subject that studies how computers simulate or implement human learning behavior to acquire new knowledge or skills, reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent. Its application is widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and example-based learning. For example, in this application, the object feature vector of the object can be generated by machine learning.

[0042] In order to better understand the scheme of the embodiments of the present application, the related terms and concepts that may be involved in the embodiments of the present application will be introduced first.

[0043] People targeting: before the advertiser launches the advertisement, the candidate people need to be circled as the targeting condition at the launch end. People targeting is the first step of launching the advertisement.

[0044] Social diffusion: when the existing interactive people of the advertisement are too few to be directly used for people targeting, the social diffusion function is needed to select more potential interactive people from the large plate people to form the targeting condition. The goal of social diffusion is to maximize the social interaction rate of the advertisement in the friend circle and the click rate of the advertisement. Here, social interaction refers to the number of likes and comments received by an advertisement in the friend circle.

[0045] Object embedding: the features of the object are generally stored in the underlying database in a discrete high-dimensional sparse manner. In order to efficiently process online services, it is necessary to map the high-dimensional sparse features of the object to a low-dimensional dense space, which is called object embedding.

[0046] BERT: English full name Bidirectional Encoder Representations from Transformers, which means Bidirectional Encoder Representations from Transformers, a language representation model.

[0047] An important component in the advertising targeting service is the crowd diffusion. When an advertiser needs to place an advertisement on a social platform, in addition to considering the click rate of the advertisement on the social platform, the advertisement also needs to consider obtaining more object interactions (likes and comments). The traditional crowd diffusion algorithm, because it is only based on the characteristics of the object itself for calculation, ignores the social interaction relationship between objects, so it cannot meet the social diffusion needs in the friend circle scene. The existing social diffusion algorithm combines the object social influence score (reflecting the potential number of social interactions that the object can bring) and the object interest score (based on the characteristics of the object itself and the characteristics of close friends) to find the target crowd. Although it can improve the click rate and interaction rate at the same time, there are two problems: on the one hand, the model training and prediction time is long. The existing algorithm needs to generate training samples for each advertiser's social diffusion task, train the object interest score model, then calculate the interest score and social influence score for the large plate object respectively, and finally select the object with the largest comprehensive score to generate the social diffusion crowd. Because each process is an offline computing task, the time for one task is more than 1 hour, and a large amount of computing resources are consumed. On the other hand, it does not consider the importance of different close friend characteristics in the object interest score. The existing algorithm directly averages the characteristics of the closest friends of the object after the model training, which will result in that the friend characteristics of different objects have no distinction, introducing additional noise to the interest score model, and reducing the accuracy of the object interest score prediction.

[0048] To solve the above problems, an embodiment of the present application proposes a fast social diffusion method based on object embedding vectors, which pre-calculates the vector characteristics of the object and its close friends, thereby omitting the time for model training for each advertiser individually, and greatly reducing the model prediction time. In the pre-computation stage, an embodiment of the present application proposes a social interaction scoring model based on a graph attention model, which can determine the importance of the friend characteristics in the object representation according to the interaction intensity of the object to each friend and the object characteristics of the object and the friend; in the prediction stage, an embodiment of the present application directly reads the object embedding information with friend characteristics from the database to form the embedding vector of the seed object, and then determines whether the object requesting the advertisement is a social diffusion object by using the threshold filtering method in the advertisement recall stage. Compared with the current crowd targeting method using the number package, the embodiment of the present application greatly reduces the time from extraction to effectiveness of crowd targeting, and because of the introduction of the attention mechanism, the embodiment of the present application can improve the click rate and interaction rate of the social diffusion crowd when placing an advertisement.

[0049] Please refer to Figure 1 , Figure 1This is a schematic diagram of the architecture of a data processing system provided in an embodiment of this application. Figure 1 As shown, the architecture of the data processing system may include a server 100 and terminal devices 101. There may be multiple terminal devices 101, and each terminal device 101 may be connected to the server 100 via a network. The network connection may be a wired connection or a wireless connection, so that each terminal device can interact with the server 100 through the network connection, and the server 100 can receive business data from each terminal device.

[0050] The terminal device 101 can be used to display business data to the target audience. This business data can specifically include the number of seed users, the magnitude of the target audience, advertising platforms, and audience expansion trends. Broadly speaking, it can include advertising data and target audience data. Advertising data includes, for example, advertising category, advertising ID, advertiser ID, ad impression rate, click-through rate, and interaction rate. Target audience data includes, for example, target audience ID, target audience tags, and target audience tags. The terminal device 101 can also store some business data, such as pulling seed users from the server 100 and storing them offline. In addition, the terminal device 101 can integrate and install a target application to manage advertising and target audience data, acting as a data management platform. This target application can also run as a webpage on the terminal device 101. The server 100 can store the business data generated by the target application and interact with each terminal device 101. For example, the target audience can upload target data to the server 100 through the target application, or download target data or advertising data from the server 100.

[0051] In the embodiments of the present application, the server 100 can filter similar objects from the big plate objects according to the seed crowd uploaded by the advertiser, and the similar objects form a diffusion crowd. The advertiser can bind the advertisement to the diffusion crowd. The system recalls the advertisements bound to the similar objects in the advertisement recall stage as the candidate set of the object, and then selects some appropriate advertisements to the object. Specifically, the server 100 runs a social diffusion model, models the task of predicting the object social diffusion score as a binary classification problem, and can comprehensively train the model by using the three parts of data, i.e., the object features of each object in the big plate object, the interaction intimacy between the objects, and the interaction information between the objects and the advertisements in the sample data. The object feature vector finally obtained is stored in the data engine. When the model is predicted, the object feature vector can not only determine the seed vector corresponding to the advertiser, but also further determine the similar objects according to the seed vector and the object feature vector, and the similar objects are used as the diffusion crowd. In the advertisement recall system, the advertisements bound to the diffusion crowd by the advertiser are recalled. In this process, the importance of friends is distinguished by the interaction intimacy, so that the weight proportions of the friend features in the representation of the object feature vector are different, so as to more accurately depict the object feature vector, so that the target diffusion crowd is quickly determined according to the object feature vector, and the data efficiency of the advertisement delivery is higher.

[0052] The above scheme is mainly used for a module of a new crowd in a data management platform for advertisement marketing service: crowd expansion. Please refer to Figure 2 , Figure 2 is an interface diagram of a crowd expansion extraction module provided by the embodiments of the present application, as shown in Figure 2 , the advertiser can browse the attributes of the uploaded seed crowd in the seed crowd 20, set the expected diffusion number in the target crowd number 21, and select whether to contain the seed crowd in the seed crowd containing 22, to determine whether to deliver the seed crowd in the final advertisement delivery. When the object expansion tendency selects social interaction, the delivery platform selects the friend circle, the background will call the above social diffusion algorithm to generate a target crowd number size crowd package and store it in the Hadoop Distributed File System (HDFS), and then the crowd package will be pushed to the delivery end of the advertisement, and used as the object customized crowd for advertisement targeting.

[0053] In actual implementation, the data processing system can be deployed based on a blockchain network, i.e., the server 100 and the object terminal 101 can be both deployed in the blockchain network, or the object terminal 101 is deployed outside the blockchain network and the server 100 is deployed in the blockchain network, etc. When the server 100 is deployed in the blockchain network and the server 100 is a server cluster or a distributed system composed of multiple physical servers, each physical server can be a node in the blockchain network. The blockchain mentioned herein is a new application mode of distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm, etc. computer technology. Its essence is a decentralized database, which is a series of data blocks associated using cryptography. In the data processing method disclosed in the present application, the data (such as the object feature vector of each object, the object feature, the interactive closeness, etc.) can be saved on the blockchain.

[0054] The terminal device in the embodiments of the present application can include a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart television, a desktop computer, a smart watch, etc. smart terminal device, but is not limited thereto.

[0055] It can be understood that the method provided by the embodiments of the present application can be executed by a computer device, which includes but is not limited to the server 100 described above. The server 100 can be a standalone physical server, or a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDNs, and basic cloud computing services such as big data and artificial intelligence platforms.

[0056] Please refer to Figure 3 , Figure 3 is a flowchart of a data processing method provided by the embodiments of the present application. The execution subject in the embodiments can be a computer device or a cluster composed of multiple computer devices. The computer device can be a terminal device or a server. Herein, the execution subject in the embodiments is taken as an example to illustrate the server. As shown in Figure 3 , the data processing method includes but is not limited to the following steps S301-S305:

[0057] Step S301, obtaining an object set.

[0058] In a possible embodiment, the object set (i.e., the user set) can include N objects (i.e., users), where N is a positive integer, each object having its own characteristics, referred to as object characteristics, which can accurately describe the object, and the data form is usually a vector. The object characteristics of an object can include multi-dimensional characteristics of the object, including gender characteristics and interest label characteristics. For example, the characteristics of an object are described as a girl, a rock music fan, and the corresponding description words or sentences of the characteristics of the object can be obtained by a corresponding method, such as word vector extraction. Each object can have a unique identifier as an object identifier, i.e., an object ID. For example, the login account of a social platform object can be used as an object ID. In summary, the object set can be an object or an application client in a platform with an interactive function and an object identifier, for example, an object registered on a social platform. The corresponding object characteristics can be analyzed through the activity log of the social platform. The content included in the object set is not limited here.

[0059] In step S302, a target object is determined from the object set, and an interactive object of the target object is obtained from the object set.

[0060] In a possible embodiment, in a colloquial sense, the object set is a large-dish object, and advertisement targeting or other services can be implemented based on the large-dish object. When the objects in the object set need to be processed, a target object (i.e., a target user) can be randomly selected from the object set for processing. When the object set is all registered objects of a social platform, there is a certain relationship between the objects, for example, the objects follow each other and are friends, and thus each object has a friend list, and each friend in the friend list is also an object in the object set. When a target object is determined from the object set, the interactive friends corresponding to the target object can be obtained as interactive objects (i.e., interactive users). Here, the interaction can be the interaction of likes and comments between the target object and the interactive object, i.e., the interactive object likes or comments on the social dynamics published by the target object on the social platform, or the target object likes or comments on the social dynamics published by the interactive object on the social platform. Since the target object can also like or comment on the social dynamics published by itself, the target object can also be regarded as an interactive object of itself, so that the interactive objects can include the target object and the friends who interact with the target object.

[0061] In step S303, the interaction intimacy between the target object and the interactive object is obtained, and the advertisement interaction behavior information between the target object and the candidate advertisement data is obtained.

[0062] In a possible embodiment, the object interacting with the social dynamic published by the target object is taken as the interaction object, and the interaction intimacy between the target object and the interaction object can be used to measure the frequency of interaction, representing the intimacy between the target object and the interaction object. For different interaction objects, i.e., different friends of the target object, the interaction intimacy can be different, and the specific expression form of the interaction intimacy is usually the size of the numerical value. The acquisition method of the interaction intimacy can be described as follows. Since the interaction object includes the target object, the corresponding interaction intimacy between the target object and the interaction object not only includes the interaction intimacy between the target object and the interaction friend, but also includes the interaction intimacy between the target object and the target object itself.

[0063] As an optional implementation, the acquisition method of the interaction intimacy between the target object and the interaction object can include the following steps. First, object interaction behavior information between the target object and the interaction object is acquired, and the interaction times between the target object and the interaction object are acquired according to the object interaction behavior information. The object interaction behavior information can include data generated by the interaction behavior of the interaction object to the target object, including the interaction times, such as the total number of likes and comments, 5 times as the interaction times. Then, the interaction time points corresponding to the object interaction behavior information are acquired, and the time difference between the interaction time points and the current time points is acquired. The object interaction behavior information can correspond to one or more interaction time points between the target object and the interaction object. The interaction time point can be the interaction date, i.e., the specific day when the interaction object has interaction behavior to the target object. Alternatively, the interaction time point can be counted in hours, minutes, or seconds, which is not limited here. The object interaction behavior information and the interaction time points (such as the interaction date) can be stored in the object's flow log, for example, on February 1, 2021, the interaction object has any one or both of the like or comment behaviors to the target object, and the background data log will record the interaction time point as February 1, 2021. The current time point can be the starting point of the interaction intimacy calculation, and the unit of the interaction time point corresponds, for example, the current time point is March 1, 2021, and if a specific period (such as every week, every month, or every day) is set for calculation, the object interaction behavior information in the unit time in the past period can be used to obtain the time difference value corresponding to the interaction times. For example, the current time point is March 1, 2021, the period is 28 days, and the interaction times are counted in unit time in days. The interaction times of the interaction time point February 1, 2021 are 5, and the corresponding time difference value is the number of days between the current time point March 1 and the interaction time point February 1, i.e., 28. Then, the time decay weight can be determined according to the time difference value, and the interaction times are weighted according to the time decay weight to obtain the interaction intimacy. The determination of the time decay weight can be according to the following expression (1)

[0064]

[0065] wherein, a is a fixed value set by man, usually 0.5, T i represents the interaction time point, T cur represents the current time point.

[0066] Using the above example, the weighting between the number of interactions and the time decay weight is 5xW t If the target object and the interaction object have only one value in a cycle, then 5xW t can be used as the interaction intimacy. Generally speaking, the object interaction behavior information in a cycle (such as a month, a week or a day) includes the number of interactions under multiple interaction time points, each interaction time point and the current time point has a time difference value, each time difference value has a time decay weight, and there is also a corresponding number of interactions. The number of interactions at multiple interaction time points and the time decay weight are weighted and summed, and then normalized to obtain the interaction intimacy. Simply put, it is to calculate the number of interactions between the objects every day, then calculate the time decay weight according to the day difference between the interaction date and the current date, and finally output the intimacy coefficient between the friends (i.e. the interaction intimacy) obtained by weighting and summing the number of interactions according to the time decay weight and then normalizing.

[0067] For example, if the target object and the interaction object have interacted for 5 days between February 1, 2021 and March 1, 2021, respectively on February 1, February 5, February 6, February 14, and February 15, and the number of interactions per day according to the order of the interaction time point is 5, 2, 3, 4, and 6, respectively, and the day difference between the interaction date and the current date is 28, 24, 23, 15, and 14, respectively, the corresponding time decay weight is e -28α , e -24α , e -23α , e -15α , e -14α , respectively. The number of interactions and the time decay weight are correspondingly weighted and summed, specifically: W t = 5e -28α + 2e -24α + 3e -23α + 4e -15α + 6e -14α , and finally the W t is normalized to obtain the interaction intimacy. Through normalization, the data is mapped to the range of 0-1, making data processing more rapid and convenient.

[0068] As an extended way, the interaction behaviors of the object to the advertisement in a unit time (such as in the past month) can be counted, and the closeness coefficient can be calculated by using the statistical data, that is, the interaction closeness is determined by taking the interaction behaviors of the friends of the target object to the target object under the advertisement, such as responding to the comments of the target object under the advertisement, or liking the advertisement together with the target object, or liking the comments of the target object as the interaction behavior data.

[0069] In addition to the interaction behaviors between the interaction object and the target object, there are also interaction behaviors between the target object and the candidate advertisement data. Unlike the interaction behaviors between the objects, the advertisement interaction behaviors include not only the likes and comments on the advertisement, but also the clicks on the advertisement. Such interaction behaviors will generate advertisement interaction behavior information, such as the behaviors of the target object to click, like and comment on the advertisement appearing on the social dynamic display page, wherein the likes and comments can be collectively referred to as interactions. The candidate advertisement data is a plurality of advertisements ready to be put on the terminal of the target object, or a plurality of advertisements ready to be put on the terminal of the object again, wherein each advertisement includes data such as advertisement identifier (i.e. advertisement ID) and advertisement exposure frequency set by the advertiser.

[0070] In step S304, according to the interaction closeness, the advertisement interaction behavior information and the object characteristics of the interaction object, the correlation weight value between the target object and the interaction object is generated.

[0071] In a possible embodiment, the specific generation process of the association weight value between the target object and the interactive object can include the following steps: first, input the interaction intimacy and the object features of the interactive object into an initial prediction model, and generate an initial association weight value between the target object and the interactive object through an attention mechanism structure in the initial prediction model. Since the interactive object can include the target object, the interactive object features include the object features of the target object and the object features of the interactive friends of the target object, and the object features of each object are low-dimensional embedding vectors (or encoding vectors), which are specifically determined by processing the high-dimensional original object features through a BERT model. This is because directly using the original features of the objects and friends to input the model for training will particularly lengthen the calculation resources and time consumption of the training, and therefore, a 16-dimensional vector is generated as the object feature by using a non-supervised learning method to perform embedding on the original object features of all objects. The BERT model is a bidirectional encoder representation from transformers algorithm based on the transformer, which is a model used for encoding tokens and sentences in the field of natural language processing. When applied to the object feature embedding scenario, each feature index in the object features of the object is regarded as a "word", and the object features of each object arranged in ascending order of index are regarded as a "sentence" input into the pre-training task of the BERT. Each feature index is a word describing the role of the object, which is a feature corresponding to the word in the label of the object, for example, the label of the object includes: female, rock and roll fan, etc. Therefore, the object features corresponding to each word describing the object are actually a feature index, which together constitute the original object features of the object. As known from the above, there is no direct context relationship between the objects like the sentences, and therefore, only the masked language model (MLM) self-supervised task can be used for pre-training to obtain the final object features, without involving specific downstream tasks of the BERT. Specifically, for the original object features of each object, 15% of the indexes can be randomly covered, and then the remaining object features of the object can be used to predict all the covered indexes. For example, the sentence "my dog is cute" can be "my dog is [masked]" after being covered by the mask. Please refer to Figure 4 , Figure 4 is a partial structure diagram of a language representation model BERT provided by an embodiment of the present application. As Figure 4As shown, unlike the original BERT, only the token embeddings are used as the first layer of the model input for each index, without using segment embeddings and position embeddings. After pre-training, the special classification embedding word [CLS] representing each sentence vector in the original BERT is used as the object feature embedding of each object.

[0072] The initial prediction model is a model using an attention mechanism, and the initial correlation weight value can be obtained by processing the input interaction intimacy and the object features obtained by the BERT model through the attention mechanism. Taking a node classification network based on attention, i.e., a graph attention network (GAT) as an example, the model can predict the importance of different adjacent nodes, has good performance, and is more robust to perturbation. The input of a single graph attention layer (GAL) is a set of node feature vectors, which corresponds to the feature vector obtained by splicing the object features of the interaction objects and the respective interaction intimacies in the embodiments of the present application. For specific processing, please refer to Figure 5 , Figure 5 is a schematic diagram of an attention mechanism structure provided by the embodiments of the present application, as shown in Figure 5 The attention mechanism is applied to the graph attention layer, and first performs self-attention processing on the set of object feature vectors of the interaction objects spliced with the interaction intimacies to obtain an attention correlation coefficient, and the specific expression is as follows formula (2) :

[0073] e ij =a(Wh i ,Wh j ),j∈N i (2)

[0074] wherein a is a mapping, h i is the object feature of the target object i, h j is the feature vector obtained by splicing the object feature of the target object i and the interaction intimacy, W is a weight matrix shared by all h i , and N i represents the set of interaction objects of the target object i. Taking each target object as a node, then the target object i is node i, and N i is the set of adjacent nodes of node i, and node i also belongs to N iPart of the target object i, i.e., the target object i can also be regarded as the interactive friend of itself. For the allocation of attention a, the embodiment of the application is allocated on the object feature vector set of the interactive friend.

[0075] The obtained attention mutual correlation coefficient e ij The initial correlation weight value is obtained through the softmax layer output, and the specific expression is as follows formula (3) :

[0076]

[0077] Among them, α ij is the initial correlation weight value between the target object i and the interactive object j, e ij is the attention coefficient processed by the self-attention mechanism, e ik is the attention coefficient between the target object i and the interactive object k.

[0078] The above initial correlation weight value is the calculation of the interactive friend weight. The embedding of the object itself (i.e., the object feature of the target object), the embedding of the friend (i.e., the object feature of the interactive object except the target object) and the intimacy (i.e., the interactive intimacy) are used as inputs, and a three-layer neural network is used to calculate the inputs h i and h j are both feature vectors composed of corresponding interactive intimacy and object features.

[0079] Then, the initial object feature vector of the target object is generated according to the initial correlation weight value. For details, please refer to Figure 6 , Figure 6 is a scene diagram for generating an object feature vector provided by the embodiment of the application, as shown in Figure 6 The original graph attention model uses a multi-head attention mechanism to model the importance of each neighbor node feature, wherein h1 is the target object feature vector generated by the BERT model by splicing the interactive intimacy between itself and itself, h1 to h6 are the interactive friend object feature vectors spliced with the intimacy between the target object and the corresponding interactive friend, and α is the importance of each interactive friend vector to the object h1 vector. Here, α is calculated using multi-head attention, and different colors are used in the figure to represent different weights, which can make the weight have better robustness. The number of heads of the multi-head attention is 3 by default. In the multi-layer graph attention layer included in the graph attention model, the input and output of each layer are a set of node feature vectors, and only the dimension of each node feature vector in the new set of node feature vectors output by each layer can not be equal to the dimension of the input node feature vector. The initial object feature vector of the target object generated here corresponds to h1' in Figure 6 , which can be stored in the data engine, and it also needs to be updated according to the update of the correlation weight value in the training process.

[0080] Then the advertisement feature vector of the candidate advertisement data is obtained, the model parameters of the initial prediction model are corrected according to the initial object feature vector, the advertisement interaction behavior information and the obtained advertisement feature vector, and a target prediction model is obtained. In the initial prediction model, the input advertisement interaction behavior information is obtained from the training samples extracted from the advertisement flow log, and the objects that click and have interaction behavior can be found in the advertisement flow of the social platform in the last day, each object constitutes a positive sample of (object ID, advertisement ID, whether to click, whether to interact), here the positive sample at least has a click or interaction on the advertisement, then other exposed advertisement objects are regarded as negative samples (that is, the objects that do not click and do not interact on the advertisement), wherein whether to click and whether to interact can use numerical 1 and 0 as sample labels, that is, 0 for no click (or interaction) and 1 for click (or interaction). Because usually the proportion of interactive objects is generally only 1% of the clicked objects, the oversampling technology can be used to expand the number of interactive samples by a ratio of 1:5, for example, when there is an interactive sample 1 (object 0074, advertisement 1135, 0, 1) in the extracted sample, 5 copies of the same interactive sample are added to the training sample, which can solve the problem of uneven distribution of data in the training process and improve the accuracy of model training.

[0081] The training samples (including positive samples and negative samples) are input into the initial prediction model. Specifically, according to the object ID and the advertisement ID included in the training samples, the corresponding initial object feature vector and the advertisement feature vector can be determined from the corresponding data engine. According to the initial object feature vector and the advertisement feature vector, the interaction prediction value of the target object for the candidate advertisement data can be determined, which represents the possibility of the target object clicking or interacting with the candidate advertisement data. The specific implementation manner can be to calculate the cosine similarity between the initial object feature vector and the advertisement feature vector, and to obtain a probability value between 0 and 1 through a softmax layer as the interaction prediction value. Alternatively, the cosine similarity can be directly used as the interaction prediction value to measure the possibility of the target object clicking and interacting with the candidate advertisement data. According to the interaction prediction value and the advertisement interaction behavior information, the prediction loss value of the initial prediction model can be determined. Since the advertisement interaction behavior information includes the aforementioned sample label, which represents whether the object has an interaction or click behavior with the advertisement, this sample label serves as a reference standard. The loss calculation can be performed according to the 0 or 1 carried by the label and the interaction prediction value to obtain the prediction loss value. The specific manner of loss calculation can be a conventional cross-entropy calculation, which will not be described here. The specific manner is not further limited. The obtained prediction loss value is back propagated to the initial prediction model. Based on the back-propagated prediction loss value, the model parameters of the initial prediction model can be corrected. When the final loss value converges, the training of the initial model is stopped, and the target prediction model is obtained. Alternatively, the interaction prediction value can be obtained by processing the object-side features and the advertisement-side features using a double-tower model. The object-side features are the initial object feature vector output by the last layer of the graph attention model, and the advertisement-side features are the advertisement feature vector of the candidate advertisement data, including the first-level category, the second-level category, and the advertiser ID. The introduction of the double-tower model can make the object feature vectors of different objects that have clicked and interacted with the same advertisement closer in the vector space, thereby enabling faster determination of the target diffusion population for advertisement distribution.

[0082] After the initial prediction model is trained to obtain the target prediction model, the association weight value between the target object and the interaction object can be generated based on the attention mechanism structure in the target prediction model. The association weight value is continuously updated based on the initial association weight value. During the training process of the initial prediction model, the model parameters are updated as the prediction loss value is back propagated, and thus the association weight value also changes.

[0083] In step S305, the object feature vector of the target object is generated according to the association weight value and the object feature of the interaction object.

[0084] In a possible embodiment, the object quantity of the interaction object has at least two, one of which is a target object, that is, at least one is an interaction friend. The specific steps of generating the object feature vector of the target object by using the association weight value and the object feature of the interaction object can include: weighting the object features of the at least two interaction objects according to the association weight value to obtain the weighted features corresponding to each interaction object respectively. As shown in the above Figure 6 The association weight value between h1 and h2 is α 12 , the association weight value and the object feature of the interaction object are weighted to obtain α 12 h2, where the object feature of the interaction object is the feature of the interaction closeness. The weighted features of different objects can be obtained by splicing the weighted values corresponding to each interaction object respectively, and the object feature vector can be obtained, and the specific expression is as follows formula (4):

[0085]

[0086] Where ∥ represents splicing of vectors, K represents the number of attention mechanisms, k represents the k-th in K, σ represents a nonlinear activation function, Wk represents the association weight value corresponding to the k-th attention mechanism, and W k represents the linear transformation of the input interaction object feature, h j represents the object feature of the interaction object j.

[0087] Since the final output of the concate operation can not be so sensitive, the weighted features of the interaction objects can also be directly averaged to obtain the object feature vector, and the specific expression is as follows:

[0088]

[0089] Where σ represents a nonlinear activation function, K represents the number of attention mechanisms, k represents the k-th in K, σ represents a nonlinear activation function, Wk represents the association weight value corresponding to the k-th attention mechanism, and W k represents the linear transformation of the input interaction object feature, h j represents the object feature of the object j.

[0090] The object feature vector generated according to the above method can be used as the final output of the graph attention model, and the final object feature vector of the object obtained after model training will also be loaded into the data engine. Optionally, the object feature vector of the target object can be used to determine the advertising data for the target object. The combination of the above-mentioned graph attention network model and the double tower model for processing data can be collectively referred to as a social diffusion model (or a social diffusion scoring model), and the logical processing algorithm thereof can be referred to as a social diffusion algorithm. Please refer to Figure 7 ,Figure 7 This is a flowchart illustrating a social diffusion algorithm provided in an embodiment of this application, such as... Figure 7 As shown, the specific process can include steps ① to ⑥. Steps ① to ③ can be performed simultaneously to save computation time, steps ① to ⑤ are performed only once a day, and step ⑥ is only invoked after the advertiser initiates a social diffusion request. Step ① extracts training samples from the ad flow, including the aforementioned positive and negative samples. The ad interaction behavior information can be obtained through the sample data. Step ② calculates the intimacy between the object and the interacting friend. For details, please refer to the relevant content in step S303. Step ③ corresponds to the object self-supervised encoding vector implemented using the BERT model in step S304. However, if computational resources are limited and it is not possible to train the BERT model to generate object embeddings every day, Principal Component Analysis (PCA) and Word Vector Mapping (word2vec) can be used to quickly train and generate object embeddings. This is because both PCA and word2vec have Spark library functions that can be quickly deployed in a cluster. Step ④ feature concatenation includes ad-side features and object-side features. Ad-side features are mainly used in the dual-tower model, while object-side features are modeled and processed using Graph Attention Network (GAT). The ad-side features use the first-level category, second-level category, and advertiser ID of the ad as features. The object-side features consist of two parts: the object feature embedding of the object itself and the feature embedding of the object's interactive friends. The features of each object's interactive friends are constructed by concatenating the friend's object feature embedding with the intimacy level between the object and the interactive friend. To improve computational efficiency, the maximum number of interactive friends considered for each object is set to 250. If there are more than 250 interactive friends, only the 250 friends with the highest intimacy level are considered as the input data for the object's interactive friend features. Furthermore, to further improve performance, the object's interactive friend features can be extended to the object's second-order interactive friend features, that is, the features of the interactive friends of the object's interactive friends are also considered. The model training process in step ⑤ can be found in steps S304 and S305; the model prediction and online ad recall in step ⑥ can be found below. Figure 8 and Figure 9 The corresponding method implementations will not be described in detail here.

[0091] The model can be used to model the task of predicting the social diffusion score of an object as a binary classification problem, i.e., judging whether the input object belongs to the diffusion object of the advertiser or can be understood as whether the object has a click or interaction behavior on the advertiser's advertisement. By adopting a centralized training method to train a unified social diffusion scoring model for all advertisers with recent demand, the time for training a separate model for each advertiser is reduced, the click sample and interaction sample are mixed as positive samples of the model, the weight of click scoring and interaction scoring in the final scoring is learned by using the existing click and interaction sample distribution of the advertiser, which avoids the suboptimal diffusion result caused by manual determination of the weight, and also realizes the effect of different weight scoring for each advertiser. At the same time, a graph attention model is introduced to model the friend intimacy and the difference between different friend object features, and an embedding is generated for each object for prediction.

[0092] It should be noted that the object feature vector of the target object can be obtained in the same way, and the advertisement for each object can be determined in the same way.

[0093] In summary, the embodiments of the present application have at least the following advantages:

[0094] By the interaction intimacy between the target object and the interaction object, the advertisement interaction behavior information, and the object features of the interaction object, the correlation weight value between the target object and the interaction object is generated, and then the object feature vector of the target object is generated according to the correlation weight value and the object features spliced with the interaction intimacy. In this process, a unified target prediction model can be trained for different advertisers, and effective computing resources are used to meet the demand of full-quantity advertisers for finding target diffusion population, and the weight of click scoring and interaction scoring of the model is trained by using the advertisement interaction behavior information, which can automatically find the optimal weight for each advertiser. In addition, the interaction intimacy between different interaction objects and the target object is added to the corresponding interaction object feature vector, and the features of different interaction friends are distinguished, the initial correlation weight value obtained by inputting the initial prediction model for training is updated constantly with the back propagation of the loss value to obtain the final correlation weight value, and the importance of different friend features to the object feature vector of the target object is distinguished according to the correlation weight value. Since the graph attention network with attention mechanism is introduced into the model, the model is more robust, the target prediction model obtained after training is more accurate in scoring the social diffusion of the object, and the object feature vector obtained finally is also more robust and the information expressed by the object feature vector is more accurate and rich, so that the click rate and interaction rate of the object on the advertisement data after the advertisement data is determined by using the object feature vector can be effectively improved.

[0095] Please refer to Figure 8 ,Figure 8 is a flowchart of a data processing method provided by an embodiment of the present application. The execution subject in the embodiment can be a computer device or a cluster composed of multiple computer devices. The computer device can be a terminal device or a server. Here, the execution subject in the embodiment is taken as an example for illustration. As shown in Figure 8 the data processing method includes but is not limited to the following steps S801-S804:

[0096] Step S801, obtaining the to-be-placed advertisement data of the advertisement data placement object, and determining the target object as a candidate placement object associated with the advertisement data placement object.

[0097] In a possible embodiment, the advertisement data placement object is an object of advertisement placement using an advertisement placement system, collectively referred to as an advertiser. Each advertiser has planned advertisements stored in the advertisement library corresponding to the advertisement placement system, and part or all of the advertisement data in the advertisement library can be prepared for placement to the object as to-be-placed advertisement data. The advertisement placement system can be a functional module applied in an application client or a separate placement platform in the application client, and the advertisement library is a data engine for storing all advertisement data of the advertiser. The object included in the seed population uploaded by the advertiser can be a candidate placement object. The object included in the seed population can also be referred to as a seed object. The seed object can have a click or interaction behavior on the advertisement or has become a conversion object of the advertisement. Accordingly, the seed population can include one or more of the clicked population, the interacted population, and the converted population, which can be used as a reference standard for population diffusion. Since the seed object also belongs to the large-dish object, the seed object can also be selected for advertisement placement when selecting the advertisement placement object (i.e., the target diffusion population). As shown in the above Figure 2 placement selection condition, the seed population 22 can be included or not included. In the embodiment of the present application, the seed population is included as an example for illustration. The target object determined from the large-dish object is the seed object that meets the screening condition after the advertiser initially places the advertisement. For the candidate placement object, the corresponding advertisement data can be specified by the advertiser or the corresponding advertisement data can have been placed to the candidate placement object, and thus the candidate placement object is associated with the advertiser.

[0098] Step S802, selecting a similar object of the candidate placement object from the object set according to the vector similarity between the object feature vector of the candidate placement object and the object feature vector of each object in the object set.

[0099] In a possible embodiment, when the advertiser uploads a seed group, the object feature vector of the seed object can be quickly retrieved by the data engine. Specifically, the object feature vector of the corresponding object can be retrieved by the object identifier of the seed object. The data engine stores the object feature vector and the object identifier trained in the foregoing embodiments. Then, average pooling is performed on all retrieved object feature vectors to obtain a seed vector of the same dimension as the object feature vector, which is used to measure the degree of similarity between the objects in the object set and the seed object. The similar objects are determined from the object set by calculating the distance between the vectors and comparing the distance value with a threshold value.

[0100] As an optional implementation, the specific steps of determining the similar objects can include: first, selecting an object subset from the object set, for example, the object set includes 1 billion objects, and the object subset includes 100,000 objects; then determining an object similarity threshold according to the vector similarity between the object feature vector of the candidate delivery object and the object feature vector of each object in the object subset; performing average on the object feature vectors of the plurality of candidate delivery objects to obtain a seed vector; if the candidate delivery object is only one, the object feature vector of the candidate delivery object is used as the seed vector; determining the object similarity threshold from the object subset according to the similarity between the seed vector and each vector in the object subset, the ratio of the target diffusion population number specified by the advertiser to the object set, and the ratio of the target diffusion population number specified by the advertiser to the object set. For example, the target diffusion population number set by the advertiser is 100 million, and the corresponding quantile ratio is 10% in the case of the object set of 1 billion. The corresponding number value of 10% in the object subset is 100,000. The seed vector and the object feature vector of each object in the object subset are arranged in descending order of similarity, and the similarity corresponding to the 100,000th object is selected as the object similarity threshold. In the object set of 1 billion objects, 100 million objects are selected according to the object similarity threshold. Finally, the objects whose vector similarity between the object feature vector in the object set and the object feature vector of the candidate delivery object is greater than the object similarity threshold are determined as the similar objects. The vector similarity between the object feature vector of each object in the object set and the seed vector is threshold truncated, that is, the object similarity threshold and the vector similarity are compared, and the objects greater than the object similarity threshold are used as the target delivery objects. The vector similarity can be obtained by calculating the inner product of the two vectors. After the similar objects are determined, the similar object IDs can be input into a file, which can be used for binding with other data to realize more functions.

[0101] It can be seen that the inner product of the seed vector and the object vector is calculated in a smaller random object set, the threshold of the corresponding quantile is calculated according to the proportion of the target diffusion population in the large plate population, the scoring (i.e. vector similarity) of the large plate object is filtered by the threshold, and the objects with the inner product greater than the threshold are retained. Thus, the threshold can be determined by sorting a smaller object subset to determine similar objects, and the target diffusion population is determined without sorting the large plate objects, which greatly saves the computing resources, reduces the computing cost, and thus speeds up the computing speed.

[0102] In step S803, the candidate delivery object and the similar object are determined as the target delivery object of the to-be-delivered advertising data.

[0103] In a possible embodiment, the to-be-delivered advertising data is the advertising delivery data specified by the advertiser for the target delivery object. The candidate delivery object, as the seed object, can have a higher conversion rate for the advertising contribution, and can bring higher benefits to the advertiser at a lower cost. Therefore, the seed object can be selected as the target delivery object together with the similar object filtered from the large plate object during the advertising delivery process. That is, in the target diffusion population set by the advertiser, the seed object can be selected from the large plate object with a high probability, so that the seed population can be included, i.e. the target diffusion population selected by the aforementioned population diffusion module includes the seed population. Of course, the target delivery object can not include the seed object, and correspondingly, the target diffusion population can not include the seed population. These target delivery objects that meet the advertising delivery conditions (i.e. the vector similarity between the seed vector is greater than the similarity threshold) belong to the diffusion population of the corresponding advertiser. For different advertisers, the diffusion population can also be different. Since the click and interaction of the object on the advertising are considered, the target diffusion population can also be referred to as a social diffusion population.

[0104] In step S804, the to-be-delivered advertising data is delivered to the device to which the target delivery object belongs.

[0105] In a possible embodiment, the to-be-delivered advertising data can be delivered to the device to which the target delivery object belongs by the advertising delivery system. The device to which the target delivery object belongs can be a terminal device such as a mobile phone, a tablet computer, a telephone watch, a desktop computer, etc. In this embodiment, because the embedding vector (i.e. object feature vector) generated by the target prediction model can be pre-stored in the data engine, the seed population can be quickly retrieved and the seed population embedding (i.e. seed vector) can be calculated. Finally, the seed population embedding can be used to find the target customer through random threshold truncation or nearest neighbor search technology to deliver the advertising, and the corresponding content of the above steps S801-S804 is Figure 7The model prediction and online recall advertisement module is triggered after the social diffusion model request initiated by the advertiser in the flowchart of the illustrated social diffusion algorithm.

[0106] The exposure of the experimental group (directly using the scheme for social diffusion) is increased by 84.5% compared with the control group (the algorithm of the current MI population diffusion module online service), the click volume is increased by 177%, the interaction volume is increased by 164%, the click rate is increased by 50%, and the average single click cost (CPC) is decreased by 11.6%.

[0107] In summary, the embodiments of the present application have at least the following advantages:

[0108] The seed vector is quickly determined from the object feature vector set stored in the data engine from the seed population uploaded by the advertiser, the similar objects of the candidate delivery objects are determined according to the similarity of the seed vector and the object feature vector of each object in the object set, and the target delivery object is determined, wherein the threshold cutting method is used to select similar objects in the object set through the threshold value determined by the object subset, avoiding the sorting of the similarity of the large plate objects, thereby effectively improving the speed of generating the target diffusion population and saving computing resources.

[0109] Please refer to Figure 9 , Figure 9 is a flowchart of a data processing method provided by an embodiment of the present application. The execution subject in the embodiment can be a computer device or a cluster composed of multiple computer devices. The computer device can be a terminal device or a server. Here, the execution subject in the embodiment is taken as an example to illustrate a server. As shown in Figure 9 , the data processing method includes but is not limited to the following steps S901-S903:

[0110] Step S901, obtaining a candidate advertisement data set of an advertisement data delivery object.

[0111] In a possible embodiment, the advertisement data managed by the advertiser (advertisement data delivery object) can be obtained from the advertisement library as a candidate advertisement data set, wherein each candidate advertisement data includes the category of the advertisement, the advertisement ID and other related data. The difference between the candidate advertisement data set and the to-be-delivered advertisement data set is that the candidate advertisement data set cannot be directly delivered to the target delivery object, and needs to be processed and screened in a series of processes before it can be delivered to the target delivery object, for example, the recall processing of the advertisement.

[0112] Step S902, establishing a binding relationship between the candidate advertisement data set and the target delivery object.

[0113] In a possible embodiment, after determining the target delivery object by truncating with the threshold value, it is also necessary to bind the candidate advertisement data set of the advertiser with the target delivery object as the advertisement specified by the advertiser for the target delivery object, and a binding relationship is established between the candidate advertisement data set of the advertiser and the target delivery object. As long as any one data is obtained, all the bound data can be pulled. The target delivery object is an object in the target diffusion group set by the advertiser, and the target diffusion group can be regarded as a tag on the target delivery object. For example, the target diffusion group tag of the advertiser A is Ad1, and the target delivery object determined according to the seed group of the advertiser A has the tag Ad1. The target diffusion group tag of the advertiser B is Ad2, and the target delivery object determined according to the seed group of the advertiser B has the tag Ad2. The target diffusion group tag corresponding to each advertiser and the ID of the target delivery object can be bound together, so that the target object in the object set can be bound with multiple target diffusion group tags corresponding to different advertisers. Therefore, the candidate advertisement data set bound to the target delivery object can also be regarded as the candidate advertisement data set bound to the target diffusion group to which the target delivery object belongs.

[0114] In step S903, when the advertisement data recall instruction is obtained, the advertisement data for delivering to the target delivery object is recalled from the candidate advertisement data set having a binding relationship with the target delivery object.

[0115] In a possible embodiment, when the object with the specified tag appears in the social platform, for example, the Ad2 object with the target diffusion group tag of the advertiser B appears when refreshing the social dynamic information, the device to which the target object belongs sends an advertisement data recall request to the corresponding server. When the server corresponding to the advertisement recall system obtains the advertisement data recall instruction, all the candidate advertisement data sets bound to the target object (that is, the candidate advertisement data sets bound to the target diffusion group) are recalled, and then the candidate advertisement data sets are subjected to rough sorting, fine sorting, strategy display, and the like, to screen the advertisement data set for delivering to the target delivery object.

[0116] As an optional mode, the seed vector can also be directly pushed to the artificial neural network (ANN) branch of the advertisement recall system as a recall vector of the bound advertisement. When the target object appears on the social platform, the ANN branch retrieves the social diffusion embedding vector of the object (i.e., the object feature vector of the target object) from the data engine, and then uses the nearest neighbor search (NN) to retrieve the most similar seed vector and the bound advertisement. Unlike the label branch, the ANN branch does not pre-record the label of the target object belonging to the target diffusion group, but judges in real time whether the target object belongs to the target diffusion group of the advertiser and whether it can carry the corresponding label. Among them, the nearest neighbor search is to find the most similar item from the database according to the similarity of the data, that is, to select the most similar recall vector from the recall vectors corresponding to multiple advertisers according to the similarity between vectors, and the recall vector is the seed vector corresponding to the seed group uploaded by the advertiser. The advertisement bound to the target diffusion group of the advertiser is the advertisement bound to the target object. Therefore, the advertisement recall system can recall the bound advertisement.

[0117] In summary, the embodiments of the present application have at least the following advantages:

[0118] Through different branches of the advertisement recall system, including the label branch and the artificial neural network branch, different processing is performed on the related data of the object, and the bound advertisement of the target diffusion group to which the object belongs is quickly recalled. Specifically, in the label branch, the bound relationship between the candidate advertisement data set and the target delivery object and the target diffusion group label carried by the target delivery object are used to recall the candidate advertisement data according to the correspondence between the label and the candidate advertisement data set. By labeling the target delivery object, the candidate advertisement set can be quickly located and recalled when the target delivery object appears. In the artificial neural network branch, the relationship between vectors is processed using the nearest neighbor technique to retrieve the candidate advertisement data set corresponding to the most similar seed object of the target delivery object for recall processing. This avoids the pre-labeling method and judges the target diffusion group label of the appearing object in real time, which is more convenient than the processing of the label branch.

[0119] Please refer to Figure 10 , Figure 10 is a structural schematic diagram of a data processing device provided by an embodiment of the present application. The data processing device can be a data processing device running in Figure 1The computer program (including program code) in the server 100 shown is an application software, for example, a data processing apparatus; the data processing apparatus can be used to execute corresponding steps in the data processing method provided in the embodiments of the present application. The data processing apparatus includes an acquisition module 1001, a determination module 1002, and a generation module 1003, wherein:

[0120] The acquisition module 1001 is configured to acquire an object set; the object set includes N objects and object features of each object in the N objects, where N is a positive integer;

[0121] The determination module 1002 is configured to determine a target object from the object set, and acquire an interactive object of the target object from the object set; the interactive object includes the target object.

[0122] The acquisition module 1001 is further configured to acquire an interactive intimacy between the target object and the interactive object, and acquire advertisement interactive behavior information between the target object and candidate advertisement data.

[0123] The generation module 1003 is configured to generate an association weight value between the target object and the interactive object according to the interactive intimacy, the advertisement interactive behavior information, and the object features of the interactive object.

[0124] The generation module 1003 is further configured to generate an object feature vector of the target object according to the association weight value and the object features of the interactive object; the object feature vector of the target object is used to determine advertisement data to be delivered to the target object.

[0125] In an embodiment, the acquisition module 1001 is specifically configured to acquire object interactive behavior information between the target object and the interactive object, acquire an interaction frequency between the target object and the interactive object according to the object interactive behavior information, acquire an interaction time point corresponding to the object interactive behavior information, acquire a time difference value between the interaction time point and a current time point, determine a time decay weight according to the time difference value, weight the interaction frequency according to the time decay weight, and obtain the interactive intimacy.

[0126] In an embodiment, the generation module 1003 is specifically configured to input the interactive intimacy and the object features of the interactive object into an initial prediction model, generate an initial association weight value between the target object and the interactive object based on an attention mechanism structure in the initial prediction model, generate an initial object feature vector of the target object according to the initial association weight value, acquire an advertisement feature vector of the candidate advertisement data, correct model parameters of the initial prediction model according to the initial object feature vector, the advertisement feature vector, and the advertisement interactive behavior information, to obtain a target prediction model, and generate the association weight value between the target object and the interactive object based on an attention mechanism structure in the target prediction model.

[0127] In an embodiment, the generating module 1003 is specifically configured to: determine an interaction prediction value of the target object for the candidate advertisement data according to the initial object feature vector and the advertisement feature vector; determine a prediction loss value of the initial prediction model according to the interaction prediction value and the advertisement interaction behavior information; and correct the model parameter of the initial prediction model based on the prediction loss value to obtain the target prediction model.

[0128] In an embodiment, the generating module 1003 is further specifically configured to: weight the object features of the at least two interaction objects according to the association weight values to obtain the weighted features corresponding to each interaction object respectively; and perform splicing processing on the weighted features corresponding to each interaction object respectively to obtain the object feature vector, or perform average processing on the weighted features corresponding to each interaction object respectively to obtain the object feature vector.

[0129] In an embodiment, the data processing apparatus further includes a selecting module 1004 and a delivering module 1005, wherein:

[0130] The obtaining module 1001 is configured to obtain the advertisement data to be delivered to the advertisement data delivery object, and determine the target object as a candidate delivery object associated with the advertisement data delivery object.

[0131] The selecting module 1004 is configured to select similar objects of the candidate delivery object from the object set according to the vector similarity between the object feature vector of the candidate delivery object and the object feature vector of each object in the object set.

[0132] The determining module 1002 is configured to determine the candidate delivery object and the similar objects as the target delivery objects of the advertisement data to be delivered.

[0133] The delivering module 1005 is configured to deliver the advertisement data to be delivered to the device to which the target delivery object belongs.

[0134] In an embodiment, the selecting module 1004 is specifically configured to: select an object subset from the object set; determine an object similarity threshold according to the vector similarity between the object feature vector of the candidate delivery object and the object feature vector of each object in the object subset; and determine the object in the object set as the similar object, if the vector similarity between the object feature vector of the object and the object feature vector of the candidate delivery object is greater than the object similarity threshold.

[0135] In an embodiment, the data processing apparatus further includes an establishing module 1006 and a recalling module 1007, wherein:

[0136] The obtaining module 1001 is configured to obtain a candidate advertisement data set of the advertisement data delivery object.

[0137] The establishing module 1006 is configured to establish a binding relationship between the candidate advertisement data set and the target delivery object.

[0138] The recall module 1007 is configured to, when the advertisement data recall instruction is acquired, recall the advertisement data for the target delivery object from the candidate advertisement data set having a binding relationship with the target delivery object.

[0139] It can be understood that the functions of the various functional modules of the data processing apparatus described in the embodiments of the present application can be specifically implemented according to the methods in the above method embodiments, and the specific implementation process can refer to the related description of the above method embodiments, which will not be repeated here.

[0140] Please refer to Figure 11 , Figure 11 is a structural schematic diagram of a computer device provided by the embodiments of the present application, which can include a processor 1101, a memory 1102, a network interface 1103 and at least one communication bus 1104. Among them, the processor 1101 for scheduling computer programs can include a central processor, a controller, a microprocessor; the memory 1102 for storing computer programs can include a high-speed random access memory, a non-volatile memory such as a magnetic disk storage device, a flash memory device; the network interface 1103 provides data communication function, and the communication bus 1104 is responsible for connecting various communication elements.

[0141] The processor 1101 can be configured to call the computer program in the memory to perform the following operations:

[0142] Obtain an object set; the object set includes N objects and object features of each object in the N objects, where N is a positive integer;

[0143] Determine a target object from the object set, and obtain an interactive object of the target object from the object set; the interactive object includes the target object;

[0144] Obtain an interactive intimacy between the target object and the interactive object, and obtain advertisement interactive behavior information between the target object and candidate advertisement data;

[0145] According to the interactive intimacy, the advertisement interactive behavior information and the object features of the interactive object, generate an association weight value between the target object and the interactive object;

[0146] According to the association weight value and the object features of the interactive object, generate an object feature vector of the target object; the object feature vector of the target object is used to determine the advertisement data for the target object.

[0147] In an embodiment, the processor 1101 is specifically configured to: obtain object interaction behavior information between the target object and the interaction object, and obtain an interaction frequency between the target object and the interaction object according to the object interaction behavior information; obtain an interaction time point corresponding to the object interaction behavior information, and obtain a time difference value between the interaction time point and a current time point; determine a time decay weight according to the time difference value, and weight the interaction frequency according to the time decay weight to obtain an interaction closeness.

[0148] In an embodiment, the processor 1101 is specifically configured to: input the interaction closeness and object features of the interaction object into an initial prediction model; generate an initial association weight value between the target object and the interaction object based on an attention mechanism structure in the initial prediction model, and generate an initial object feature vector of the target object according to the initial association weight value; obtain an advertisement feature vector of the candidate advertisement data; correct model parameters of the initial prediction model according to the initial object feature vector, the advertisement feature vector and advertisement interaction behavior information to obtain a target prediction model; and generate an association weight value between the target object and the interaction object based on an attention mechanism structure in the target prediction model.

[0149] In an embodiment, the processor 1101 is specifically configured to: determine an interaction prediction value of the target object for the candidate advertisement data according to the initial object feature vector and the advertisement feature vector; determine a prediction loss value of the initial prediction model according to the interaction prediction value and the advertisement interaction behavior information; correct the model parameters of the initial prediction model based on the prediction loss value to obtain the target prediction model.

[0150] In an embodiment, the processor 1101 is specifically configured to: weight object features of at least two interaction objects according to the association weight value to obtain weighted features corresponding to each interaction object respectively; perform splicing processing on the weighted features corresponding to each interaction object respectively to obtain the object feature vector; or perform average processing on the weighted features corresponding to each interaction object respectively to obtain the object feature vector.

[0151] In an embodiment, the processor 1101 is further configured to: obtain to-be-delivered advertisement data of an advertisement data delivery object, and determine the target object as a candidate delivery object associated with the advertisement data delivery object; select a similar object of the candidate delivery object from the object set according to a vector similarity between an object feature vector of the candidate delivery object and an object feature vector of each object in the object set; determine the candidate delivery object and the similar object as target delivery objects of the to-be-delivered advertisement data; and deliver the to-be-delivered advertisement data to a device to which the target delivery objects belong.

[0152] In an embodiment, the processor 1101 is specifically configured to: select an object subset from the object set; determine an object similarity threshold according to a vector similarity between an object feature vector of the candidate advertising object and an object feature vector of each object in the object subset; and determine an object in the object set as a similar object if the vector similarity between the object feature vector of the object and the object feature vector of the candidate advertising object is greater than the object similarity threshold.

[0153] In an embodiment, the processor 1101 is further configured to: obtain a candidate advertising data set of the advertising data; establish a binding relationship between the candidate advertising data set and the target advertising object; and when an advertising data recall instruction is obtained, recall the advertising data of the target advertising object from the candidate advertising data set that has the binding relationship with the target advertising object.

[0154] In a specific implementation, the processor 1101, the memory 1102, and the network interface 1103 described in the embodiments of the present application can perform the implementation mode of the computer device described in the data processing method provided by the embodiments of the present application, and can also perform the implementation mode and beneficial effects described in the data processing apparatus provided by the embodiments of the present application, which will not be described here.

[0155] The embodiments of the present application also provide a computer readable storage medium, which stores instructions. When the instructions are run on a computer, the computer executes the data processing method according to the embodiments of the present application. The specific implementation can refer to the foregoing description, which will not be described here.

[0156] The embodiments of the present application also provide a computer program product or a computer program, which includes computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the data processing method according to the embodiments of the present application. The specific implementation can refer to the foregoing description, which will not be described here.

[0157] It should be noted that, for each of the foregoing method embodiments, in order to simply describe, each is described as a combination of a series of actions, but those skilled in the art should know that the present application is not limited to the order of the actions described, because according to the present application, some steps can be performed in other order or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0158] Those skilled in the art can understand that all or part of the steps of various methods in the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer readable storage medium, which can include a flash disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0159] The above merely provides a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present application, and these modifications or replacements shall be encompassed within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A data processing method, characterized by, The method comprises the following steps: acquiring an object set; the object set comprises N objects and object features of each of the N objects, wherein N is a positive integer; determining a target object from the object set and acquiring an interactive object of the target object from the object set; the interactive object comprises the target object; acquiring an interactive closeness between the target object and the interactive object, and acquiring advertisement interactive behavior information between the target object and candidate advertisement data; generating an association weight value between the target object and the interactive object according to the interactive closeness, the advertisement interactive behavior information and object features of the interactive object; generating an object feature vector of the target object according to the association weight value and the object features of the interactive object; the object feature vector of the target object is used to determine advertisement data for delivery to the target object; wherein the generating of the association weight value between the target object and the interactive object according to the interactive closeness, the advertisement interactive behavior information and the object features of the interactive object comprises: inputting the interactive closeness and the object features of the interactive object into an initial prediction model; generating an initial association weight value between the target object and the interactive object based on an attention mechanism structure in the initial prediction model, and generating an initial object feature vector of the target object according to the initial association weight value; acquiring an advertisement feature vector of the candidate advertisement data, and correcting model parameters of the initial prediction model according to the initial object feature vector, the advertisement feature vector and the advertisement interactive behavior information to obtain a target prediction model; generating the association weight value between the target object and the interactive object based on the attention mechanism structure in the target prediction model.

2. The method of claim 1, wherein, The acquiring of the interactive closeness between the target object and the interactive object comprises: acquiring object interactive behavior information between the target object and the interactive object, and acquiring an interactive frequency between the target object and the interactive object according to the object interactive behavior information; acquiring an interactive time point corresponding to the object interactive behavior information, and acquiring a time difference value between the interactive time point and a current time point; determining a time decay weight according to the time difference value, weighting the interactive frequency according to the time decay weight to obtain the interactive closeness.

3. The method of claim 1, wherein, The correcting of the model parameters of the initial prediction model according to the initial object feature vector, the advertisement feature vector and the advertisement interactive behavior information to obtain the target prediction model comprises: determining an interactive prediction value of the target object for the candidate advertisement data according to the initial object feature vector and the advertisement feature vector; determining a prediction loss value of the initial prediction model according to the interactive prediction value and the advertisement interactive behavior information; correcting the model parameters of the initial prediction model based on the prediction loss value to obtain the target prediction model.

4. The method of claim 1, wherein, The number of the interactive objects is at least two; the generating of the object feature vector of the target object according to the association weight value and the object features of the interactive object comprises: According to the association weight value, object features of at least two interactive objects are weighted to obtain a weighted feature corresponding to each interactive object respectively; The weighted features corresponding to each interactive object respectively are spliced to obtain the object feature vector; or The weighted features corresponding to each interactive object respectively are averaged to obtain the object feature vector.

5. The method of claim 1, wherein, The method further comprises: obtaining advertisement data to be put into an advertisement data putting object, determining the target object as a candidate putting object associated with the advertisement data putting object; selecting a similar object of the candidate putting object from the object set according to the vector similarity between the object feature vector of the candidate putting object and the object feature vector of each object in the object set; determining the candidate putting object and the similar object as target putting objects of the advertisement data to be put; putting the advertisement data to be put into a device to which the target putting object belongs.

6. The method of claim 5, wherein, The method further comprises: selecting an object subset from the object set; determining an object similarity threshold according to the vector similarity between the object feature vector of the candidate putting object and the object feature vector of each object in the object subset; determining an object in the object set as the similar object if the vector similarity between the object feature vector of the object and the object feature vector of the candidate putting object is greater than the object similarity threshold.

7. The method of claim 5, wherein, The method further comprises: obtaining a candidate advertisement data set of the advertisement data putting object; establishing a binding relationship between the candidate advertisement data set and the target putting object; when an advertisement data recall instruction is obtained, recalling advertisement data to be put into the target putting object from the candidate advertisement data set having the binding relationship with the target putting object.

8. A computer device, comprising: comprises: a network interface, a processor and a memory, the network interface, the processor and the memory are connected, the network interface is used to provide data communication function, the memory is used to store computer program, the processor is used to call the computer program to execute the data processing method as claimed in any one of claims 1-7.

9. A storage medium, characterized by The storage medium stores a computer program, the computer program comprises program instructions, and the program instructions are executed by the processor to execute the data processing method as claimed in any one of claims 1-7.

Images